\nplan:
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fcb632ed386a1d566da64abfc367d102b%2FSelection_999(2255).png?generation=1687603062931131&alt=media\")
to be updated …
\nplan:
\n
baseline results with public open-source yolov8,v7
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F949a2ff9f4d403746ad3270b64efd1c6%2FSelection_999(2310).png?generation=1687862996888778&alt=media\")
updated results for mmdet3 - cascade-rcnn-resnext101
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F5d74e46fd86c67c9248a9c9f26cf280a%2FSelection_999(2693).png?generation=1689439570270454&alt=media\")
notebook:
\nhttps://www.kaggle.com/code/hengck23/lb4-09-baseline-yolov7?scriptVersionId=134762462
to be updated …
\nplan:
\n
baseline results with public open-source yolov8,v7
\n
updated results for mmdet3 - cascade-rcnn-resnext101
\n
notebook:
\nhttps://www.kaggle.com/code/hengck23/lb4-09-baseline-yolov7?scriptVersionId=134762462
i study all images carefully and find that the header image below is not in the downloaded data ….
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F0c06793c94adfb8533f018bac9efc778%2FSelection_999(2337).png?generation=1688009293141835&alt=media\")
haha maybe it is from other related famous dataset
", "rawMarkdown": "haha maybe it is from other related famous dataset", "votes": 1, "replies": [ { "id": 2326195, "postDate": "2023-07-02T01:40:21.783Z", "content": "maybe it is wsi5 … i see if i can match the banner header in one of the 30 wsi in previous compeition.
", "rawMarkdown": "maybe it is wsi5 ... i see if i can match the banner header in one of the 30 wsi in previous compeition.\n" } ] } ] }, { "id": 2315708, "postDate": "2023-06-24T10:39:16.490Z", "content": "dataset insight
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F6e8220459e613e8c48e579bdc9ea4510%2FSelection_999(2256).png?generation=1687603154319438&alt=media\")
you may want to think about why \"unsure\" mask is zero in some dataset set
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F0ffd5f146f4eac980fe706b9a8750635%2FSelection_999(2257).png?generation=1687603178583579&alt=media\")
where wsi5?
\n\n", "rawMarkdown": "where wsi5?\n> dataset insight\n> \n> .png?generation=1687603154319438&alt=media)\n\n" } ] }, { "id": 2317363, "postDate": "2023-06-25T16:35:46.400Z", "content": "dataset insight
\n\n
a well known trick to improve small object instance segmentation
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fb914b0a3acb2efdb8ccbef2f187ea15c%2FSelection_999(2279).png?generation=1687710944748712&alt=media\")
So I would assume model 1 can segment better than model 2. Unet can perform segmentation better than Instance segmentation model, is it right?
", "rawMarkdown": "So I would assume model 1 can segment better than model 2. Unet can perform segmentation better than Instance segmentation model, is it right?", "replies": [ { "id": 2317773, "postDate": "2023-06-26T01:40:13.917Z", "content": "yes. this information (model 1 prediction) is transmitted to model 2.
\nif model2 is already high resolution, then you do not need two models. just include semantic segmentation as aux loss in model 2.
\nor sometimes people just scale up (e.g. 1024x1024) input to model2
", "rawMarkdown": "yes. this information (model 1 prediction) is transmitted to model 2.\n\n\n---\n\nif model2 is already high resolution, then you do not need two models. just include semantic segmentation as aux loss in model 2. \n\nor sometimes people just scale up (e.g. 1024x1024) input to model2", "votes": 2 }, { "id": 2318413, "postDate": "2023-06-26T10:52:34.320Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F72b14eed8e0aafebd403de6fc60fc40c%2FSelection_999(2287).png?generation=1687776706448050&alt=media\")
you can run unet are multi scale (e.g. at tile assembled image to capture large structure)
", "rawMarkdown": ".png?generation=1687776706448050&alt=media)\n\nyou can run unet are multi scale (e.g. at tile assembled image to capture large structure)", "votes": 1 }, { "id": 2318444, "postDate": "2023-06-26T11:20:05.970Z", "content": "\ncheck previous kaggle competition:
\ne.g. https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/discussion/298081
it feed mask-rcnn output into unet instead
", "rawMarkdown": "@ptran1203 \n\ncheck previous kaggle competition:\ne.g. https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/discussion/298081\n\nit feed mask-rcnn output into unet instead", "votes": 1 }, { "id": 2318468, "postDate": "2023-06-26T11:45:14.797Z", "content": "Thank for sharing, it's great to see you in this competition
", "rawMarkdown": "Thank for sharing, it's great to see you in this competition " } ] } ] }, { "id": 2316503, "postDate": "2023-06-25T03:20:22.667Z", "content": "backup of results
\n25-jun : lb 0.409
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F73111e631c913c67bc8f119d5dcda1e1%2FSelection_999(2265).png?generation=1687663082457630&alt=media\")
Can you public to inference segment yolov8 esemble ?, I have try but failed.
", "rawMarkdown": "Can you public to inference segment yolov8 esemble ?, I have try but failed. ", "replies": [ { "id": 2318422, "postDate": "2023-06-26T11:05:46.447Z", "content": "check past kaggle competition,like
\nSartorius - Cell Instance Segmentation
\nhttps://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/overview
https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion/notebook
\n\"Ensembling masks from different models with customed Weighted Boxes Fusion\"
Hi, could you pls give some explanations about the 'train/val' column? Since I am trying to use my custom dataset(randomly samples from dataset1 as test set, while the rest is split into train&val) on the yolov7-seg model with the same parameter setting up as you do, but I couldn't get an obvious jump inprovement between the inference with dilation and the baseline(inference without dilation). I am quite confused that whether there is a problem of my dataset to lead the weird result.
", "rawMarkdown": "Hi, could you pls give some explanations about the 'train/val' column? Since I am trying to use my custom dataset(randomly samples from dataset1 as test set, while the rest is split into train&val) on the yolov7-seg model with the same parameter setting up as you do, but I couldn't get an obvious jump inprovement between the inference with dilation and the baseline(inference without dilation). I am quite confused that whether there is a problem of my dataset to lead the weird result." } ] }, { "id": 2323447, "postDate": "2023-06-30T00:39:10.130Z", "content": "matched!!!!
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F919c2ba83bfed839851be831516d4fdf%2FSelection_999(2378).png?generation=1688085546916487&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F6e45324d6f55d8e844a70becb5a5514f%2FSelection_999(2382).png?generation=1688086199323777&alt=media\")
Is it mean WSI 3 matched with last competition dataset?
", "rawMarkdown": "Is it mean WSI 3 matched with last competition dataset?" } ] }, { "id": 2322478, "postDate": "2023-06-29T10:01:12.283Z", "content": "will context (assembled tiles) helps?
\ndataset1 blood vessel: red=sure, green=unsure
\ndataset2 blood vessel: yellow=unknown
from the tile boundary, we see both transition from yellow to red and yellow to green.
\nhence i can conclude that unknown = sure + unsure
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F92a9ebba56311b522983c2a3532c0602%2FSelection_999(2354).png?generation=1688033996218654&alt=media\")
looks like polygon masks were made directly on tiles instead of on the whole image.
", "rawMarkdown": "looks like polygon masks were made directly on tiles instead of on the whole image." } ] }, { "id": 2316842, "postDate": "2023-06-25T09:18:12.423Z", "content": "you probably have to write your data loader
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F195bbfcf854d5c8a84fbbdac68cc781b%2FSelection_999(2276).png?generation=1687684689402462&alt=media\")
In order to remove the edge effects, when inferring yolo learned with an image size of 512, we obtained the margin from the adjacent patch and predicted the test as size 640 → discarded predictions other than the central 512^512 prediction, resulting in LB487- > dropped to 484.
\n(ref: https://www.kaggle.com/competitions/hubmap-kidney-segmentation/discussion/238013)
is the training also using adjacent patch? if not, you will see performance drop
\n(this can be proven by setting train = 512, infer = 512,640, etc .. the scale is different and cause performance drop)
a better solution:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff91c6a4c1b6325059bb17922bc8e8187%2FSelection_999(2281).png?generation=1687724192099686&alt=media\")
i design this to help human in the loop annotation but it cannot be used.
", "rawMarkdown": "a better solution:\n.png?generation=1687724192099686&alt=media)\n\ni design this to help human in the loop annotation but it cannot be used.", "votes": 1 }, { "id": 2317321, "postDate": "2023-06-25T15:46:41.923Z", "content": "how to convert dataset.2(non-expert) to dataset.1 (expert) labels:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fdefcf83b7dbf1ce49f7f9eb6ace8eaaf%2FSelection_999(2278).png?generation=1687707996541137&alt=media\")
i just realize that we can include \"unsure\" detection in submission …. they just needs to be lower rank (lower score) than the \"sure\" detection.
\nHence we are actually having an instance segmentation + ranking problem.
\nhence dataset.1 has rank ground truth whereas dataset.2 has not.
you can also use multi-ask SAM for ranking:
\ninput = prompt + image
\noutput = refine mask + rank score
Whether the semi-supervised algorithm is effective in this competition？
", "rawMarkdown": "Whether the semi-supervised algorithm is effective in this competition？" }, { "id": 2321180, "postDate": "2023-06-28T11:02:12.543Z", "content": "Do you mean add \"unsure\" prediction into submission? or determine each \"unsure\" is blood_vessel and its probability?
", "rawMarkdown": "Do you mean add \"unsure\" prediction into submission? or determine each \"unsure\" is blood_vessel and its probability?" } ] }, { "id": 2320642, "postDate": "2023-06-28T02:00:36.493Z", "content": "I have the same though. However, what I concern is how to merge pseudo prediction with sparse-level annotation in DS2
", "rawMarkdown": "I have the same though. However, what I concern is how to merge pseudo prediction with sparse-level annotation in DS2", "votes": 1 } ] }, { "id": 2316053, "postDate": "2023-06-24T16:06:36.073Z", "content": "how to add and transpose x,y augmention (equivalent to rotate90) to yolo
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F092537e9a4327eaf931031a2d444028c%2FSelection_999(2261).png?generation=1687622753749040&alt=media\")
then verify batch the train output
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F53f5b691089d069d937b7c6ccd2823cc%2FSelection_999(2260).png?generation=1687622794058384&alt=media\")
how to use dataset1. to relabel dataset.2 . split into sure and unsure(or keep as unknown)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd25df8269eec787c610ffc61b3ccfe9b%2FSelection_999(2390).png?generation=1688098738371886&alt=media\")
Maybe yolov8 is better than yolov7 through local visual judgment, but i have a question, how to install ultralytics package?
\nMy approach:
\n*. I executeed the following command line locally to generate zip file , and uploaded to Kaggle
\n!pip install ultralytics -i https://pypi.douban.com/simple\n !pip show ultralytics\n ! wheels\n% wheels\n !pip wheel ultralytics\n % ../\n !zip -r ultralytics.zip wheels\n*. When ran the following command in kaggle, i encountered errors, the main error type is incompatible device type.
\n\n!pip install --no-index --no-deps /kaggle/input/yolov8-dependency/ultralytics2/wheels2/*.whl\nI will appreciate it if u can help me.
", "rawMarkdown": "Maybe yolov8 is better than yolov7 through local visual judgment, but i have a question, how to install ultralytics package?\nMy approach:\n*. I executeed the following command line locally to generate zip file , and uploaded to Kaggle\n```sh\n# local\n!pip install ultralytics -i https://pypi.douban.com/simple\n !pip show ultralytics\n !mkdir wheels\n%cd wheels\n !pip wheel ultralytics\n %cd ../\n !zip -r ultralytics.zip wheels\n```\n*. When ran the following command in kaggle, i encountered errors, the main error type is incompatible device type.\n```sh\n#in Kaggle\n!pip install --no-index --no-deps /kaggle/input/yolov8-dependency/ultralytics2/wheels2/*.whl\n```\n\nI will appreciate it if u can help me.", "votes": 2, "replies": [ { "id": 2321363, "postDate": "2023-06-28T13:52:07.183Z", "content": "just like this
\n!cp -r /kaggle/input/yolov8/ultralytics-main/ /kaggle/working/ultralytics-main
\n!pip install /kaggle/working/ultralytics-main --no-index --find-links=/kaggle/input/yolov8/ultralytics-main/
Thanks a lot!
", "rawMarkdown": "Thanks a lot!" } ] }, { "id": 2321366, "postDate": "2023-06-28T13:54:39.337Z", "content": "ultralytics-offline
\npython v3.7
\nwish i have the time its seems so interesting… at least i can read and learn
Wow! thank you , it works
", "rawMarkdown": "Wow! thank you , it works " } ] }, { "id": 2322023, "postDate": "2023-06-29T03:30:16.437Z", "content": "actually it is possible just to use
\nsys..('< to ultralytics>')\nI dint understand what you meant by 1 class : blood_vessel + not glumerulus
", "rawMarkdown": "I dint understand what you meant by 1 class : blood_vessel + not glumerulus", "votes": 2, "replies": [ { "id": 2320999, "postDate": "2023-06-28T08:04:46.840Z", "content": "so do I. I guess he use only \"blood_vessel\", no \"glomerulus\" + no \"unsure\".
\nCould you verify what I guess? @hengck23
I think it is removing the glumerulus' from the blood vessel masks.
\nSomething like: blood_vessel_mask = blood_vessel_mask * ( 1 - glumerulus_mask ) for all glumerulus and blood vessell masks
Its inference code contains' unsure ', and I believe its training phase should include two other types besides blood vessels
", "rawMarkdown": "Its inference code contains' unsure ', and I believe its training phase should include two other types besides blood vessels" } ] } ] }, { "id": 2319631, "postDate": "2023-06-27T08:36:33.317Z", "content": "the 72e40acccadf.tif test image is actually from dataset.2, wsi3
\nso if we jsut triain on dataset.2, wsi3,4 and make a submission, the lb score (which is dataset.1, wsi3,4) tells us how good are the annotations!
", "rawMarkdown": "the 72e40acccadf.tif test image is actually from dataset.2, wsi3\n\nso if we jsut triain on dataset.2, wsi3,4 and make a submission, the lb score (which is dataset.1, wsi3,4) tells us how good are the annotations!\n\n", "votes": 2 }, { "id": 2316510, "postDate": "2023-06-25T03:33:08.037Z", "content": "SAM variants:
\n[1] fast SAM : https://github.com/CASIA-IVA-Lab/FastSAM
\n[2] high quality SAM : https://github.com/SysCV/SAM-HQ
\n[3] domain adaption: https://tianrun-chen.github.io/SAM-Adaptor/static/pdfs/Adaptor.pdf
https://github.com/Hedlen/awesome-segment-anything
", "rawMarkdown": "SAM variants:\n[1] fast SAM : https://github.com/CASIA-IVA-Lab/FastSAM\n[2] high quality SAM : https://github.com/SysCV/SAM-HQ\n[3] domain adaption: https://tianrun-chen.github.io/SAM-Adaptor/static/pdfs/Adaptor.pdf\n\nhttps://github.com/Hedlen/awesome-segment-anything", "votes": 2 }, { "id": 2351334, "postDate": "2023-07-20T04:11:41.580Z", "content": "How are you using these high dimension I. e. 1440x1440 are you stitching the patchea and patchifying them?
", "rawMarkdown": "How are you using these high dimension I. e. 1440x1440 are you stitching the patchea and patchifying them? " }, { "id": 2347493, "postDate": "2023-07-17T02:47:21.147Z", "content": "Thanks for the excellent discussion points. I was interested in SAM myself and looked into the paper and code. Wanted to confirm my understanding, I think the decoder of SAM creates a channel for each instance of a mask. Do you specify an arbitrary number of masks for your decoder?
", "rawMarkdown": "Thanks for the excellent discussion points. I was interested in SAM myself and looked into the paper and code. Wanted to confirm my understanding, I think the decoder of SAM creates a channel for each instance of a mask. Do you specify an arbitrary number of masks for your decoder?" }, { "id": 2346735, "postDate": "2023-07-16T13:29:43.807Z", "content": "MMDET2 and MMDET3 appear to be very close at a resolution of 1440.
\nmy mmdet2 settings:
\n nms iou 0.5 ,
\n score_th:0.05 , (but 0.01 got the same score)
\n ep 18/20
@atom1231atom1231
\nmore experimental details
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F04e1d72c87a624b94d9205d0ae397c55%2FSelection_999(2701).png?generation=1689553581351781&alt=media\")
may i know what does each column represent for segm_mAP_copypaste?
", "rawMarkdown": "may i know what does each column represent for segm_mAP_copypaste?" }, { "id": 2347641, "postDate": "2023-07-17T05:35:07.040Z", "content": "Evaluate annotation*segm*\nDONE (=1.74s).\nAccumulating evaluation results.\nDONE (=0.23s).\n Average Precision (AP) @[ =0.50:0.95 | area= all | =100 ] = 0.301\n Average Precision (AP) @[ =0.50 | area= all | =1000 ] = 0.459\n Average Precision (AP) @[ =0.75 | area= all | =1000 ] = 0.297\n Average Precision (AP) @[ =0.50:0.95 | area= small | =1000 ] = 0.081\n Average Precision (AP) @[ =0.50:0.95 | =medium | =1000 ] = 0.221\n Average Precision (AP) @[ =0.50:0.95 | area= large | =1000 ] = 0.347\n Average Recall (AR) @[ =0.50:0.95 | area= all | =100 ] = 0.358\n Average Recall (AR) @[ =0.50:0.95 | area= all | =300 ] = 0.358\n Average Recall (AR) @[ =0.50:0.95 | area= all | =1000 ] = 0.358\n Average Recall (AR) @[ =0.50:0.95 | area= small | =1000 ] = 0.135\n Average Recall (AR) @[ =0.50:0.95 | =medium | =1000 ] = 0.289\n Average Recall (AR) @[ =0.50:0.95 | area= large | =1000 ] = 0.380\n07/15 18:04:34 - mmengine - - \n+--------------+-------+--------+--------+-------+-------+-------+\n| category | mAP | mAP_50 | mAP_75 | mAP_s | mAP_m | mAP_l |\n+--------------+-------+--------+--------+-------+-------+-------+\n| glomerulus | 0.651 | 0.798 | 0.718 | 0.0 | 0.394 | 0.786 |\n| blood_vessel | 0.253 | 0.578 | 0.172 | 0.244 | 0.269 | 0.256 |\n| unsure | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |\n+--------------+-------+--------+--------+-------+-------+-------+\n07/15 18:04:34 - mmengine - - segm_mAP_copypaste: 0.301 0.459 0.297 0.081 0.221 0.347\nI see, in what way did you do your split?
", "rawMarkdown": "I see, in what way did you do your split?" }, { "id": 2351586, "postDate": "2023-07-20T08:50:42.080Z", "content": "You can try to train only the blood vessel category, which is my experience
", "rawMarkdown": "You can try to train only the blood vessel category, which is my experience" } ] } ] }, { "id": 2341457, "postDate": "2023-07-12T06:04:15.753Z", "content": "Thank you so much for providing this experiment result. It really help me hack this competition and building the first step.
", "rawMarkdown": "Thank you so much for providing this experiment result. It really help me hack this competition and building the first step." }, { "id": 2322910, "postDate": "2023-06-29T15:11:44.520Z", "content": "i think i know how hidden slide 5 would look like
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F6de2e7ba3fe9da1e08163c89a30a27c0%2FSelection_999(2359).png?generation=1688051493403063&alt=media\")
check also paper https://www.biorxiv.org/content/10.1101/2021.11.09.467810v1.full.pdf
", "rawMarkdown": "i think i know how hidden slide 5 would look like\n\n.png?generation=1688051493403063&alt=media)\n\ncheck also paper https://www.biorxiv.org/content/10.1101/2021.11.09.467810v1.full.pdf", "replies": [ { "id": 2322954, "postDate": "2023-06-29T15:34:39.777Z", "content": "you can download about 30 WSI from humap website . each file is about 0.5 to 1.0 GB.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff378bb89a4eac8806ef61ed08bbcb2d7%2FSelection_999(2364).png?generation=1688052842335199&alt=media\")
\n \n \nIs this a leak (same donor, though WSI maybe different)?
", "rawMarkdown": "you can download about 30 WSI from humap website . each file is about 0.5 to 1.0 GB.\n\n\n.png?generation=1688052842335199&alt=media)\n\n```\n\"You may find resources from the previous HuBMAP competitions useful as well:\n\n HuBMAP: Hacking the Kidney\n HuBMAP + HPA: Hacking the Human Body\"\n```\n\nIs this a leak (same donor, though WSI maybe different)?\n", "replies": [ { "id": 2323038, "postDate": "2023-06-29T16:27:56.520Z", "content": "But they're not labelled though, right?
", "rawMarkdown": "But they're not labelled though, right? " } ] } ] }, { "id": 2322114, "postDate": "2023-06-29T04:50:38.320Z", "content": "some of the ground truth are long and elongated.
\nThis papers talks about how to set the parameters (flow) of cellpose for these kind of labels
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ffb03bb9cbfe7c42361958a154f686a11%2FSelection_999(2338).png?generation=1688014170049301&alt=media\")
[1] Omnipose: a high-precision morphology-independent solution for bacterial cell segmentation
\nhttps://www.nature.com/articles/s41592-022-01639-4
[2] Misic, a general deep learning-based method for the high-throughput cell segmentation of complex bacterial communities
\nhttps://elifesciences.org/articles/65151
nice paper
\n\"Learning Melanocytic Proliferation Segmentation in Histopathology Images from Imperfect Annotations\"
\nhttps://openaccess.thecvf.com/content/CVPR2021W/CVMI/papers/Liu_Learning_Melanocytic_Proliferation_Segmentation_in_Histopathology_Images_From_Imperfect_Annotations_CVPRW_2021_paper.pdf
\"To alleviate the cost of data annotation, we leverage a sparse annotation pipeline. Our model can be trained on sparse and noisy labels and achieves state-of-the-art performance in identifying melanocytic proliferations, producing a segmentation with Dice score 0.719, mIOU 0.740 and overall pixel accuracy 0.927.\"
", "rawMarkdown": "nice paper\n\"Learning Melanocytic Proliferation Segmentation in Histopathology Images from Imperfect Annotations\"\nhttps://openaccess.thecvf.com/content/CVPR2021W/CVMI/papers/Liu_Learning_Melanocytic_Proliferation_Segmentation_in_Histopathology_Images_From_Imperfect_Annotations_CVPRW_2021_paper.pdf\n\n\"To alleviate the cost of data annotation, we leverage a sparse annotation pipeline. Our model can be trained on sparse and noisy labels and achieves state-of-the-art performance in identifying melanocytic proliferations, producing a segmentation with Dice score 0.719, mIOU 0.740 and overall pixel accuracy 0.927.\"\n" }, { "id": 2321635, "postDate": "2023-06-28T18:06:39.133Z", "content": "Hi, do you have any inference example with yolov8?, DetectMultiBackend seems to not exist or maybe it's different in this version.
\nThanks in advance!
", "rawMarkdown": "Hi, do you have any inference example with yolov8?, DetectMultiBackend seems to not exist or maybe it's different in this version.\n\nThanks in advance!" }, { "id": 2321251, "postDate": "2023-06-28T12:04:56.350Z", "content": "Thank you for sharing your valuable information. I'm a beginner, but I'm trying various things to improve my score. I was able to notice in this thread that EPOCH is missing 10.
", "rawMarkdown": "Thank you for sharing your valuable information. I'm a beginner, but I'm trying various things to improve my score. I was able to notice in this thread that EPOCH is missing 10." }, { "id": 2320086, "postDate": "2023-06-27T13:35:22.847Z", "content": "Hi, which version of yolov8 are you using?
\nThanks!
", "rawMarkdown": "Hi, which version of yolov8 are you using?\n\nThanks!" }, { "id": 2318753, "postDate": "2023-06-26T15:01:14.967Z", "content": "https://blog.csdn.net/yangyu0515/article/details/130410948
\nthis gives very detailed explanation on the structure of SAM mask decoder, so that you can code from scratch or modify it
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff5fe04c5f4bc290b26e45d78a195ef10%2FSelection_999(2291).png?generation=1687792868562312&alt=media\")
combining all ideas so far
\nit is note that we can finetune SAM image encoder with MAE (auto masked encoder) using self supervsied learning for
\nunlabelled train set and external data (and even online with test data)
\"To deal with these issues, 'segmenteverygrain' relies on a Unet-style, patch-based convolutional neural network to create a first-pass segmentation which is then used as a set of prompts for the SAM-based segmentation. Some of the grains will be missed with this approach, but the segmentations that are created tend to be of high quality.\"
\nhttps://github.com/zsylvester/segmenteverygrain
", "rawMarkdown": "\"To deal with these issues, 'segmenteverygrain' relies on a Unet-style, patch-based convolutional neural network to create a first-pass segmentation which is then used as a set of prompts for the SAM-based segmentation. Some of the grains will be missed with this approach, but the segmentations that are created tend to be of high quality.\"\n\nhttps://github.com/zsylvester/segmenteverygrain" }, { "id": 2316920, "postDate": "2023-06-25T10:13:33.310Z", "content": "How can we evaluate mAP@0.6 instead of @0.5-0.95 or @0.5 in YOLO
", "rawMarkdown": "How can we evaluate mAP@0.6 instead of @0.5-0.95 or @0.5 in YOLO", "replies": [ { "id": 2317016, "postDate": "2023-06-25T11:43:57.463Z", "content": "yolov8:
\nultralytics-main/ultralytics/yolo/utils/metrics.py
model_file = \n model = YOLO(model_file)\n\n metrics = model.val() \n (,metrics.box.) \n (,metrics.box.map50) \n (,metrics.box.map75) \n (,metrics.box.maps) \n (,metrics.seg.) \n (,metrics.seg.map50) \n (,metrics.seg.map75) \n (,metrics.seg.maps) \n\n\n (,metrics.seg.all_ap) \n (SimpleClass):\n\n () -> \n .p = [] \n .r = [] \n .f1 = [] \n .all_ap = [] \n .ap_class_index = [] \n .nc = \n ...\n\n \n ():\n \n .all_ap[, ].mean() len(.all_ap) \n\n \n ():\n \n .all_ap[, ].mean() len(.all_ap) \nrelated paper and external data:
\n[1] Omni-Seg: A Single Dynamic Network for Multi-label Renal Pathology Image Segmentation using Partially Labeled Data
\n(glomerular tuft, glomerular unit, proximal tubular, distal tubular, peritubular capillaries, and arteries)
https://openreview.net/forum?id=v-z4Zxkt9Ex
\nhttps://github.com/ddrrnn123/Omni-Seg
[2] Development and evaluation of deep learning-based segmentation of histologic structures in the kidney cortex with multiple histologic stains
\n( arteries/arterioles, peritubular capillaries (PTCs))
https://www.kidney-international.org/article/S0085-2538(20)30962-5/fulltext
\n[3] AutoSAM: Adapting SAM to Medical Images by Overloading the Prompt Encoder
\nhttps://arxiv.org/pdf/2306.06370.pdf
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F077a2c2c057741a4b8bd002dfeb94716%2FSelection_999(2289).png?generation=1687781056015573&alt=media\")
to understand labelling, google for \"labeled kidney histology diagram\"
", "rawMarkdown": "related paper and external data:\n[1] Omni-Seg: A Single Dynamic Network for Multi-label Renal Pathology Image Segmentation using Partially Labeled Data\n(glomerular tuft, glomerular unit, proximal tubular, distal tubular, peritubular capillaries, and arteries)\n\nhttps://openreview.net/forum?id=v-z4Zxkt9Ex\nhttps://github.com/ddrrnn123/Omni-Seg\n\n[2] Development and evaluation of deep learning-based segmentation of histologic structures in the kidney cortex with multiple histologic stains\n( arteries/arterioles, peritubular capillaries (PTCs))\n\nhttps://www.kidney-international.org/article/S0085-2538(20)30962-5/fulltext\n\n\n[3] AutoSAM: Adapting SAM to Medical Images by Overloading the Prompt Encoder\nhttps://arxiv.org/pdf/2306.06370.pdf\n\n.png?generation=1687781056015573&alt=media)\n\n---\n\nto understand labelling, google for \"labeled kidney histology diagram\"" }, { "id": 2316702, "postDate": "2023-06-25T07:13:12.100Z", "content": "@hengck23 I saw you also applied dilation to mask, is it really safe to achieve high score in PB? doesn't make sense to me.
", "rawMarkdown": "@hengck23 I saw you also applied dilation to mask, is it really safe to achieve high score in PB? doesn't make sense to me." }, { "id": 2316513, "postDate": "2023-06-25T03:38:10.993Z", "content": "\n
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Faa2097cef3b01e06db20143a345f79c8%2FSelection_999(2268).png?generation=1687664596481492&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fdcb347e6e94f6523022498657c6827d1%2FSelection_999(2267).png?generation=1687664609272796&alt=media\")
thanks a lot for the notebook i ll be used this as baseline mine seems so bad …
\ni dont think you rallowed to human annotation if i understood correctly
thanks for the link. i have missed that message.
", "rawMarkdown": "thanks for the link. i have missed that message.", "votes": 1 } ] } ] }, { "id": 2352158, "postDate": "2023-07-20T17:07:57.600Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2320990, "postDate": "2023-06-28T07:55:27.597Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2318032, "postDate": "2023-06-26T06:43:20.483Z", "rawMarkdown": "", "votes": -11, "isDeleted": true } ], "comments": [ { "id": 2322020, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-29T03:28:15.567000", "content": "i study all images carefully and find that the header image below is not in the downloaded data ….
\n
haha maybe it is from other related famous dataset
", "votes": 1, "replies": [ { "id": 2326195, "author_name": "hengck23", "author_url": "", "post_date": "2023-07-02T01:40:21.783000", "content": "maybe it is wsi5 … i see if i can match the banner header in one of the 30 wsi in previous compeition.
", "votes": 0, "replies": [] } ] } ] }, { "id": 2315708, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-24T10:39:16.490000", "content": "dataset insight
\n
you may want to think about why \"unsure\" mask is zero in some dataset set
\n
where wsi5?
\n\n", "votes": 0, "replies": [] } ] }, { "id": 2317363, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-25T16:35:46.400000", "content": "dataset insight
\n\n
a well known trick to improve small object instance segmentation
\n
So I would assume model 1 can segment better than model 2. Unet can perform segmentation better than Instance segmentation model, is it right?
", "votes": 0, "replies": [ { "id": 2317773, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-26T01:40:13.917000", "content": "yes. this information (model 1 prediction) is transmitted to model 2.
\nif model2 is already high resolution, then you do not need two models. just include semantic segmentation as aux loss in model 2.
\nor sometimes people just scale up (e.g. 1024x1024) input to model2
", "votes": 2, "replies": [] }, { "id": 2318413, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-26T10:52:34.320000", "content": "
you can run unet are multi scale (e.g. at tile assembled image to capture large structure)
", "votes": 1, "replies": [] }, { "id": 2318444, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-26T11:20:05.970000", "content": "\ncheck previous kaggle competition:
\ne.g. https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/discussion/298081
it feed mask-rcnn output into unet instead
", "votes": 1, "replies": [] }, { "id": 2318468, "author_name": "Phat Tran", "author_url": "", "post_date": "2023-06-26T11:45:14.797000", "content": "Thank for sharing, it's great to see you in this competition
", "votes": 0, "replies": [] } ] } ] }, { "id": 2316503, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-25T03:20:22.667000", "content": "backup of results
\n25-jun : lb 0.409
\n
Can you public to inference segment yolov8 esemble ?, I have try but failed.
", "votes": 0, "replies": [ { "id": 2318422, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-26T11:05:46.447000", "content": "check past kaggle competition,like
\nSartorius - Cell Instance Segmentation
\nhttps://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/overview
https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion/notebook
\n\"Ensembling masks from different models with customed Weighted Boxes Fusion\"
Hi, could you pls give some explanations about the 'train/val' column? Since I am trying to use my custom dataset(randomly samples from dataset1 as test set, while the rest is split into train&val) on the yolov7-seg model with the same parameter setting up as you do, but I couldn't get an obvious jump inprovement between the inference with dilation and the baseline(inference without dilation). I am quite confused that whether there is a problem of my dataset to lead the weird result.
", "votes": 0, "replies": [] } ] }, { "id": 2323447, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-30T00:39:10.130000", "content": "matched!!!!
\n
Is it mean WSI 3 matched with last competition dataset?
", "votes": 0, "replies": [] } ] }, { "id": 2322478, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-29T10:01:12.283000", "content": "will context (assembled tiles) helps?
\ndataset1 blood vessel: red=sure, green=unsure
\ndataset2 blood vessel: yellow=unknown
from the tile boundary, we see both transition from yellow to red and yellow to green.
\nhence i can conclude that unknown = sure + unsure
looks like polygon masks were made directly on tiles instead of on the whole image.
", "votes": 0, "replies": [] } ] }, { "id": 2316842, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-25T09:18:12.423000", "content": "you probably have to write your data loader
\n
In order to remove the edge effects, when inferring yolo learned with an image size of 512, we obtained the margin from the adjacent patch and predicted the test as size 640 → discarded predictions other than the central 512^512 prediction, resulting in LB487- > dropped to 484.
\n(ref: https://www.kaggle.com/competitions/hubmap-kidney-segmentation/discussion/238013)
is the training also using adjacent patch? if not, you will see performance drop
\n(this can be proven by setting train = 512, infer = 512,640, etc .. the scale is different and cause performance drop)
a better solution:
\n
i design this to help human in the loop annotation but it cannot be used.
", "votes": 1, "replies": [] }, { "id": 2317321, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-25T15:46:41.923000", "content": "how to convert dataset.2(non-expert) to dataset.1 (expert) labels:
\n
i just realize that we can include \"unsure\" detection in submission …. they just needs to be lower rank (lower score) than the \"sure\" detection.
\nHence we are actually having an instance segmentation + ranking problem.
\nhence dataset.1 has rank ground truth whereas dataset.2 has not.
you can also use multi-ask SAM for ranking:
\ninput = prompt + image
\noutput = refine mask + rank score
Whether the semi-supervised algorithm is effective in this competition？
", "votes": 0, "replies": [] }, { "id": 2321180, "author_name": "阳光开朗大男孩", "author_url": "", "post_date": "2023-06-28T11:02:12.543000", "content": "Do you mean add \"unsure\" prediction into submission? or determine each \"unsure\" is blood_vessel and its probability?
", "votes": 0, "replies": [] } ] }, { "id": 2320642, "author_name": "Liam Nguyen", "author_url": "", "post_date": "2023-06-28T02:00:36.493000", "content": "I have the same though. However, what I concern is how to merge pseudo prediction with sparse-level annotation in DS2
", "votes": 1, "replies": [] } ] }, { "id": 2316053, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-24T16:06:36.073000", "content": "how to add and transpose x,y augmention (equivalent to rotate90) to yolo
\n
then verify batch the train output
\n
how to use dataset1. to relabel dataset.2 . split into sure and unsure(or keep as unknown)
\n
Maybe yolov8 is better than yolov7 through local visual judgment, but i have a question, how to install ultralytics package?
\nMy approach:
\n*. I executeed the following command line locally to generate zip file , and uploaded to Kaggle
\n!pip install ultralytics -i https://pypi.douban.com/simple\n !pip show ultralytics\n ! wheels\n% wheels\n !pip wheel ultralytics\n % ../\n !zip -r ultralytics.zip wheels\n*. When ran the following command in kaggle, i encountered errors, the main error type is incompatible device type.
\n\n!pip install --no-index --no-deps /kaggle/input/yolov8-dependency/ultralytics2/wheels2/*.whl\nI will appreciate it if u can help me.
", "votes": 2, "replies": [ { "id": 2321363, "author_name": "Ksana", "author_url": "", "post_date": "2023-06-28T13:52:07.183000", "content": "just like this
\n!cp -r /kaggle/input/yolov8/ultralytics-main/ /kaggle/working/ultralytics-main
\n!pip install /kaggle/working/ultralytics-main --no-index --find-links=/kaggle/input/yolov8/ultralytics-main/
Thanks a lot!
", "votes": 0, "replies": [] } ] }, { "id": 2321366, "author_name": "Séraphin Lampion", "author_url": "", "post_date": "2023-06-28T13:54:39.337000", "content": "ultralytics-offline
\npython v3.7
\nwish i have the time its seems so interesting… at least i can read and learn
Wow! thank you , it works
", "votes": 0, "replies": [] } ] }, { "id": 2322023, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-29T03:30:16.437000", "content": "actually it is possible just to use
\nsys..('< to ultralytics>')\nI dint understand what you meant by 1 class : blood_vessel + not glumerulus
", "votes": 2, "replies": [ { "id": 2320999, "author_name": "Dive Deeper", "author_url": "", "post_date": "2023-06-28T08:04:46.840000", "content": "so do I. I guess he use only \"blood_vessel\", no \"glomerulus\" + no \"unsure\".
\nCould you verify what I guess? @hengck23
I think it is removing the glumerulus' from the blood vessel masks.
\nSomething like: blood_vessel_mask = blood_vessel_mask * ( 1 - glumerulus_mask ) for all glumerulus and blood vessell masks
Its inference code contains' unsure ', and I believe its training phase should include two other types besides blood vessels
", "votes": 0, "replies": [] } ] } ] }, { "id": 2319631, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-27T08:36:33.317000", "content": "the 72e40acccadf.tif test image is actually from dataset.2, wsi3
\nso if we jsut triain on dataset.2, wsi3,4 and make a submission, the lb score (which is dataset.1, wsi3,4) tells us how good are the annotations!
", "votes": 2, "replies": [] }, { "id": 2316510, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-25T03:33:08.037000", "content": "SAM variants:
\n[1] fast SAM : https://github.com/CASIA-IVA-Lab/FastSAM
\n[2] high quality SAM : https://github.com/SysCV/SAM-HQ
\n[3] domain adaption: https://tianrun-chen.github.io/SAM-Adaptor/static/pdfs/Adaptor.pdf
https://github.com/Hedlen/awesome-segment-anything
", "votes": 2, "replies": [] }, { "id": 2351334, "author_name": "Bibhabasu Mohapatra", "author_url": "", "post_date": "2023-07-20T04:11:41.580000", "content": "How are you using these high dimension I. e. 1440x1440 are you stitching the patchea and patchifying them?
", "votes": 0, "replies": [] }, { "id": 2347493, "author_name": "Michael Bolton", "author_url": "", "post_date": "2023-07-17T02:47:21.147000", "content": "Thanks for the excellent discussion points. I was interested in SAM myself and looked into the paper and code. Wanted to confirm my understanding, I think the decoder of SAM creates a channel for each instance of a mask. Do you specify an arbitrary number of masks for your decoder?
", "votes": 0, "replies": [] }, { "id": 2346735, "author_name": "atom1231", "author_url": "", "post_date": "2023-07-16T13:29:43.807000", "content": "MMDET2 and MMDET3 appear to be very close at a resolution of 1440.
\nmy mmdet2 settings:
\n nms iou 0.5 ,
\n score_th:0.05 , (but 0.01 got the same score)
\n ep 18/20
@atom1231atom1231
\nmore experimental details
may i know what does each column represent for segm_mAP_copypaste?
", "votes": 0, "replies": [] }, { "id": 2347641, "author_name": "hengck23", "author_url": "", "post_date": "2023-07-17T05:35:07.040000", "content": "Evaluate annotation*segm*\nDONE (=1.74s).\nAccumulating evaluation results.\nDONE (=0.23s).\n Average Precision (AP) @[ =0.50:0.95 | area= all | =100 ] = 0.301\n Average Precision (AP) @[ =0.50 | area= all | =1000 ] = 0.459\n Average Precision (AP) @[ =0.75 | area= all | =1000 ] = 0.297\n Average Precision (AP) @[ =0.50:0.95 | area= small | =1000 ] = 0.081\n Average Precision (AP) @[ =0.50:0.95 | =medium | =1000 ] = 0.221\n Average Precision (AP) @[ =0.50:0.95 | area= large | =1000 ] = 0.347\n Average Recall (AR) @[ =0.50:0.95 | area= all | =100 ] = 0.358\n Average Recall (AR) @[ =0.50:0.95 | area= all | =300 ] = 0.358\n Average Recall (AR) @[ =0.50:0.95 | area= all | =1000 ] = 0.358\n Average Recall (AR) @[ =0.50:0.95 | area= small | =1000 ] = 0.135\n Average Recall (AR) @[ =0.50:0.95 | =medium | =1000 ] = 0.289\n Average Recall (AR) @[ =0.50:0.95 | area= large | =1000 ] = 0.380\n07/15 18:04:34 - mmengine - - \n+--------------+-------+--------+--------+-------+-------+-------+\n| category | mAP | mAP_50 | mAP_75 | mAP_s | mAP_m | mAP_l |\n+--------------+-------+--------+--------+-------+-------+-------+\n| glomerulus | 0.651 | 0.798 | 0.718 | 0.0 | 0.394 | 0.786 |\n| blood_vessel | 0.253 | 0.578 | 0.172 | 0.244 | 0.269 | 0.256 |\n| unsure | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |\n+--------------+-------+--------+--------+-------+-------+-------+\n07/15 18:04:34 - mmengine - - segm_mAP_copypaste: 0.301 0.459 0.297 0.081 0.221 0.347\nI see, in what way did you do your split?
", "votes": 0, "replies": [] }, { "id": 2351586, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-20T08:50:42.080000", "content": "You can try to train only the blood vessel category, which is my experience
", "votes": 0, "replies": [] } ] } ] }, { "id": 2341457, "author_name": "WillinKaggle", "author_url": "", "post_date": "2023-07-12T06:04:15.753000", "content": "Thank you so much for providing this experiment result. It really help me hack this competition and building the first step.
", "votes": 0, "replies": [] }, { "id": 2322910, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-29T15:11:44.520000", "content": "i think i know how hidden slide 5 would look like
\n
check also paper https://www.biorxiv.org/content/10.1101/2021.11.09.467810v1.full.pdf
", "votes": 0, "replies": [ { "id": 2322954, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-29T15:34:39.777000", "content": "you can download about 30 WSI from humap website . each file is about 0.5 to 1.0 GB.
\n
\n \n \nIs this a leak (same donor, though WSI maybe different)?
", "votes": 0, "replies": [ { "id": 2323038, "author_name": "Harry4463", "author_url": "", "post_date": "2023-06-29T16:27:56.520000", "content": "But they're not labelled though, right?
", "votes": 0, "replies": [] } ] } ] }, { "id": 2322114, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-29T04:50:38.320000", "content": "some of the ground truth are long and elongated.
\nThis papers talks about how to set the parameters (flow) of cellpose for these kind of labels
[1] Omnipose: a high-precision morphology-independent solution for bacterial cell segmentation
\nhttps://www.nature.com/articles/s41592-022-01639-4
[2] Misic, a general deep learning-based method for the high-throughput cell segmentation of complex bacterial communities
\nhttps://elifesciences.org/articles/65151
nice paper
\n\"Learning Melanocytic Proliferation Segmentation in Histopathology Images from Imperfect Annotations\"
\nhttps://openaccess.thecvf.com/content/CVPR2021W/CVMI/papers/Liu_Learning_Melanocytic_Proliferation_Segmentation_in_Histopathology_Images_From_Imperfect_Annotations_CVPRW_2021_paper.pdf
\"To alleviate the cost of data annotation, we leverage a sparse annotation pipeline. Our model can be trained on sparse and noisy labels and achieves state-of-the-art performance in identifying melanocytic proliferations, producing a segmentation with Dice score 0.719, mIOU 0.740 and overall pixel accuracy 0.927.\"
", "votes": 0, "replies": [] }, { "id": 2321635, "author_name": "Pablo Larrosa", "author_url": "", "post_date": "2023-06-28T18:06:39.133000", "content": "Hi, do you have any inference example with yolov8?, DetectMultiBackend seems to not exist or maybe it's different in this version.
\nThanks in advance!
", "votes": 0, "replies": [] }, { "id": 2321251, "author_name": "mh", "author_url": "", "post_date": "2023-06-28T12:04:56.350000", "content": "Thank you for sharing your valuable information. I'm a beginner, but I'm trying various things to improve my score. I was able to notice in this thread that EPOCH is missing 10.
", "votes": 0, "replies": [] }, { "id": 2320086, "author_name": "Pablo Larrosa", "author_url": "", "post_date": "2023-06-27T13:35:22.847000", "content": "Hi, which version of yolov8 are you using?
\nThanks!
", "votes": 0, "replies": [] }, { "id": 2318753, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-26T15:01:14.967000", "content": "https://blog.csdn.net/yangyu0515/article/details/130410948
\nthis gives very detailed explanation on the structure of SAM mask decoder, so that you can code from scratch or modify it
combining all ideas so far
\nit is note that we can finetune SAM image encoder with MAE (auto masked encoder) using self supervsied learning for
\nunlabelled train set and external data (and even online with test data)
\"To deal with these issues, 'segmenteverygrain' relies on a Unet-style, patch-based convolutional neural network to create a first-pass segmentation which is then used as a set of prompts for the SAM-based segmentation. Some of the grains will be missed with this approach, but the segmentations that are created tend to be of high quality.\"
\nhttps://github.com/zsylvester/segmenteverygrain
", "votes": 0, "replies": [] }, { "id": 2316920, "author_name": "Bibhabasu Mohapatra", "author_url": "", "post_date": "2023-06-25T10:13:33.310000", "content": "How can we evaluate mAP@0.6 instead of @0.5-0.95 or @0.5 in YOLO
", "votes": 0, "replies": [ { "id": 2317016, "author_name": "hengck23", "author_url": "", "post_date": "2023-06-25T11:43:57.463000", "content": "yolov8:
\nultralytics-main/ultralytics/yolo/utils/metrics.py
model_file = \n model = YOLO(model_file)\n\n metrics = model.val() \n (,metrics.box.) \n (,metrics.box.map50) \n (,metrics.box.map75) \n (,metrics.box.maps) \n (,metrics.seg.) \n (,metrics.seg.map50) \n (,metrics.seg.map75) \n (,metrics.seg.maps) \n\n\n (,metrics.seg.all_ap) \n (SimpleClass):\n\n () -> \n .p = [] \n .r = [] \n .f1 = [] \n .all_ap = [] \n .ap_class_index = [] \n .nc = \n ...\n\n \n ():\n \n .all_ap[, ].mean() len(.all_ap) \n\n \n ():\n \n .all_ap[, ].mean() len(.all_ap) \nrelated paper and external data:
\n[1] Omni-Seg: A Single Dynamic Network for Multi-label Renal Pathology Image Segmentation using Partially Labeled Data
\n(glomerular tuft, glomerular unit, proximal tubular, distal tubular, peritubular capillaries, and arteries)
https://openreview.net/forum?id=v-z4Zxkt9Ex
\nhttps://github.com/ddrrnn123/Omni-Seg
[2] Development and evaluation of deep learning-based segmentation of histologic structures in the kidney cortex with multiple histologic stains
\n( arteries/arterioles, peritubular capillaries (PTCs))
https://www.kidney-international.org/article/S0085-2538(20)30962-5/fulltext
\n[3] AutoSAM: Adapting SAM to Medical Images by Overloading the Prompt Encoder
\nhttps://arxiv.org/pdf/2306.06370.pdf
to understand labelling, google for \"labeled kidney histology diagram\"
", "votes": 0, "replies": [] }, { "id": 2316702, "author_name": "", "author_url": "", "post_date": "2023-06-25T07:13:12.100000", "content": "", "votes": 0, "replies": [] }, { "id": 2316513, "author_name": "", "author_url": "", "post_date": "2023-06-25T03:38:10.993000", "content": "", "votes": 0, "replies": [ { "id": 2316525, "author_name": "", "author_url": "", "post_date": "2023-06-25T04:09:26.677000", "content": "", "votes": 1, "replies": [ { "id": 2316530, "author_name": "", "author_url": "", "post_date": "2023-06-25T04:15:22.547000", "content": "", "votes": 1, "replies": [] } ] } ] }, { "id": 2352158, "author_name": "", "author_url": "", "post_date": "2023-07-20T17:07:57.600000", "content": "", "votes": 0, "replies": [] }, { "id": 2320990, "author_name": "", "author_url": "", "post_date": "2023-06-28T07:55:27.597000", "content": "", "votes": 0, "replies": [] }, { "id": 2318032, "author_name": "", "author_url": "", "post_date": "2023-06-26T06:43:20.483000", "content": "", "votes": -11, "replies": [] } ], "raw_markdown_by_id": { "2315705": "to be updated ...\nplan:\n.png?generation=1687603062931131&alt=media)\n\nbaseline results with public open-source yolov8,v7 \n\n .png?generation=1687862996888778&alt=media)\n\n\nupdated results for mmdet3 - cascade-rcnn-resnext101\n\n.png?generation=1689439570270454&alt=media)\n\nnotebook:\nhttps://www.kaggle.com/code/hengck23/lb4-09-baseline-yolov7?scriptVersionId=134762462\n\n", "2322020": "i study all images carefully and find that the header image below is not in the downloaded data ....\n.png?generation=1688009293141835&alt=media)", "2315708": "dataset insight\n\n.png?generation=1687603154319438&alt=media)", "2317363": "a well known trick to improve small object instance segmentation\n\n.png?generation=1687710944748712&alt=media)", "2316503": "backup of results\n\n25-jun : lb 0.409\n.png?generation=1687663082457630&alt=media)", "2323447": "matched!!!!\n\n.png?generation=1688085546916487&alt=media)\n\n\n.png?generation=1688086199323777&alt=media)", "2322478": "will context (assembled tiles) helps?\n\ndataset1 blood vessel: red=sure, green=unsure\ndataset2 blood vessel: yellow=unknown\n\nfrom the tile boundary, we see both transition from yellow to red and yellow to green.\nhence i can conclude that unknown = sure + unsure\n\n---\n\n.png?generation=1688033996218654&alt=media)", "2316842": "you probably have to write your data loader\n.png?generation=1687684689402462&alt=media)", "2317623": "a better solution:\n.png?generation=1687724192099686&alt=media)\n\ni design this to help human in the loop annotation but it cannot be used.", "2317321": "how to convert dataset.2(non-expert) to dataset.1 (expert) labels: \n\n.png?generation=1687707996541137&alt=media)", "2316053": "how to add and transpose x,y augmention (equivalent to rotate90) to yolo\n\n.png?generation=1687622753749040&alt=media)\n\nthen verify batch the train output\n.png?generation=1687622794058384&alt=media)\n", "2323602": "how to use dataset1. to relabel dataset.2 . split into sure and unsure(or keep as unknown)\n\n.png?generation=1688098738371886&alt=media)\n\n", "2321355": "Maybe yolov8 is better than yolov7 through local visual judgment, but i have a question, how to install ultralytics package?\nMy approach:\n*. I executeed the following command line locally to generate zip file , and uploaded to Kaggle\n```sh\n# local\n!pip install ultralytics -i https://pypi.douban.com/simple\n !pip show ultralytics\n !mkdir wheels\n%cd wheels\n !pip wheel ultralytics\n %cd ../\n !zip -r ultralytics.zip wheels\n```\n*. When ran the following command in kaggle, i encountered errors, the main error type is incompatible device type.\n```sh\n#in Kaggle\n!pip install --no-index --no-deps /kaggle/input/yolov8-dependency/ultralytics2/wheels2/*.whl\n```\n\nI will appreciate it if u can help me.", "2320945": "I dint understand what you meant by 1 class : blood_vessel + not glumerulus", "2319631": "the 72e40acccadf.tif test image is actually from dataset.2, wsi3\n\nso if we jsut triain on dataset.2, wsi3,4 and make a submission, the lb score (which is dataset.1, wsi3,4) tells us how good are the annotations!\n\n", "2316510": "SAM variants:\n[1] fast SAM : https://github.com/CASIA-IVA-Lab/FastSAM\n[2] high quality SAM : https://github.com/SysCV/SAM-HQ\n[3] domain adaption: https://tianrun-chen.github.io/SAM-Adaptor/static/pdfs/Adaptor.pdf\n\nhttps://github.com/Hedlen/awesome-segment-anything", "2351334": "How are you using these high dimension I. e. 1440x1440 are you stitching the patchea and patchifying them? ", "2347493": "Thanks for the excellent discussion points. I was interested in SAM myself and looked into the paper and code. Wanted to confirm my understanding, I think the decoder of SAM creates a channel for each instance of a mask. Do you specify an arbitrary number of masks for your decoder?", "2346735": "MMDET2 and MMDET3 appear to be very close at a resolution of 1440.\nmy mmdet2 settings: \n nms iou 0.5 , \n score_th:0.05 , (but 0.01 got the same score)\n ep 18/20", "2341457": "Thank you so much for providing this experiment result. It really help me hack this competition and building the first step.", "2322910": "i think i know how hidden slide 5 would look like\n\n.png?generation=1688051493403063&alt=media)\n\ncheck also paper https://www.biorxiv.org/content/10.1101/2021.11.09.467810v1.full.pdf", "2322114": "some of the ground truth are long and elongated.\nThis papers talks about how to set the parameters (flow) of cellpose for these kind of labels\n\n.png?generation=1688014170049301&alt=media)\n\n\n[1] Omnipose: a high-precision morphology-independent solution for bacterial cell segmentation\nhttps://www.nature.com/articles/s41592-022-01639-4\n\n[2] Misic, a general deep learning-based method for the high-throughput cell segmentation of complex bacterial communities\nhttps://elifesciences.org/articles/65151", "2322096": "nice paper\n\"Learning Melanocytic Proliferation Segmentation in Histopathology Images from Imperfect Annotations\"\nhttps://openaccess.thecvf.com/content/CVPR2021W/CVMI/papers/Liu_Learning_Melanocytic_Proliferation_Segmentation_in_Histopathology_Images_From_Imperfect_Annotations_CVPRW_2021_paper.pdf\n\n\"To alleviate the cost of data annotation, we leverage a sparse annotation pipeline. Our model can be trained on sparse and noisy labels and achieves state-of-the-art performance in identifying melanocytic proliferations, producing a segmentation with Dice score 0.719, mIOU 0.740 and overall pixel accuracy 0.927.\"\n", "2321635": "Hi, do you have any inference example with yolov8?, DetectMultiBackend seems to not exist or maybe it's different in this version.\n\nThanks in advance!", "2321251": "Thank you for sharing your valuable information. I'm a beginner, but I'm trying various things to improve my score. I was able to notice in this thread that EPOCH is missing 10.", "2320086": "Hi, which version of yolov8 are you using?\n\nThanks!", "2318753": "https://blog.csdn.net/yangyu0515/article/details/130410948\nthis gives very detailed explanation on the structure of SAM mask decoder, so that you can code from scratch or modify it ", "2318644": "\"To deal with these issues, 'segmenteverygrain' relies on a Unet-style, patch-based convolutional neural network to create a first-pass segmentation which is then used as a set of prompts for the SAM-based segmentation. Some of the grains will be missed with this approach, but the segmentations that are created tend to be of high quality.\"\n\nhttps://github.com/zsylvester/segmenteverygrain", "2316920": "How can we evaluate mAP@0.6 instead of @0.5-0.95 or @0.5 in YOLO", "2316785": "related paper and external data:\n[1] Omni-Seg: A Single Dynamic Network for Multi-label Renal Pathology Image Segmentation using Partially Labeled Data\n(glomerular tuft, glomerular unit, proximal tubular, distal tubular, peritubular capillaries, and arteries)\n\nhttps://openreview.net/forum?id=v-z4Zxkt9Ex\nhttps://github.com/ddrrnn123/Omni-Seg\n\n[2] Development and evaluation of deep learning-based segmentation of histologic structures in the kidney cortex with multiple histologic stains\n( arteries/arterioles, peritubular capillaries (PTCs))\n\nhttps://www.kidney-international.org/article/S0085-2538(20)30962-5/fulltext\n\n\n[3] AutoSAM: Adapting SAM to Medical Images by Overloading the Prompt Encoder\nhttps://arxiv.org/pdf/2306.06370.pdf\n\n.png?generation=1687781056015573&alt=media)\n\n---\n\nto understand labelling, google for \"labeled kidney histology diagram\"", "2316702": "@hengck23 I saw you also applied dilation to mask, is it really safe to achieve high score in PB? doesn't make sense to me.", "2316513": "~~looking for team members!~~\n\n~~looking for medical professionals, etc who can hand label dataset3 and other external data. ~~\n~~I want to try human in the loop feedback learning like in SAM~~\n\n.png?generation=1687664596481492&alt=media)\n\n.png?generation=1687664609272796&alt=media)", "2352158": "", "2320990": "", "2318032": "" } }