![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fdf86112888d37848ceae2b55c73cd66b%2FSelection_840.png?generation=1761586949785654&alt=media\")
1. End-to-end network
\n
1. End-to-end network
\n
host tries to make my life difficult 😭
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F25b75b54af519f8d7102964190b0ad88%2FSelection_881.png?generation=1762159108120057&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fbc7632d5f5158c1ee9f578006c92d01c%2FSelection_882.png?generation=1762159120714331&alt=media\")
😂 oh multi pages, maybe we need to get the keypoints for foreground page first. By the way, are the green points produced by your segmentation model?
", "rawMarkdown": "😂 oh multi pages, maybe we need to get the keypoints for foreground page first. By the way, are the green points produced by your segmentation model?", "replies": [ { "id": 3310643, "postDate": "2025-11-03T10:25:53.807Z", "content": "all results are raw results from the segmentation model without post processing yet. i am trying to simplify the steps, i can add a foreground object detection … but this add another step.
", "rawMarkdown": "all results are raw results from the segmentation model without post processing yet. i am trying to simplify the steps, i can add a foreground object detection ... but this add another step. " }, { "id": 3318741, "postDate": "2025-11-11T10:55:08.587Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3318742, "postDate": "2025-11-11T10:55:36.927Z", "content": "I see the foreground grid points are not connected to the main grid points, wouldn't removing the connected component with the least grid points work in your case?
", "rawMarkdown": "I see the foreground grid points are not connected to the main grid points, wouldn't removing the connected component with the least grid points work in your case?" } ] } ] }, { "id": 3308257, "postDate": "2025-10-29T00:57:11.983Z", "content": "my unwarping experiment is a success ?\npreliminary results, code to come …
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F76c18fbec58b000264bb0cd7afeeaa73%2Fecg_animation.gif?generation=1761699428005816&alt=media\")
2. Tricks and Pitfalls
\n[truth label]
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F688c36048de7be9e70ee9c9481b87a2c%2FSelection_841.png?generation=1761587201775600&alt=media\")
[predict label]
\n[context window]
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fdf0f74afcbc7dc497d9dd84267b58c8a%2FSelection_839.png?generation=1761587567631704&alt=media\")
[missing prediction]
\n[dataset infor]
\n\"Can we expect the public/private test to have a uniform distribution of the 9 augmentation types? Or is the test assembled using only type 1 like the two examples? Or was the test even constructed using augmented samples not present in the training set, such as types 2, 7, 8?\"
\nresults on validation (tarined with 100 image id over various types)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F36e485684941449476ff27a83cedd2bc%2FSelection_925.png?generation=1762538571608129&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fffc1544183682457858f0b3884ce302c%2FSelection_924.png?generation=1762537902540596&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F819d9653a120fc592fdc6fac4ec1ca3a%2FSelection_923.png?generation=1762537922062908&alt=media\")
my segmentation net learns local waveform and not the row label of lead signals … i need to redesign my network …
", "rawMarkdown": "results on validation (tarined with 100 image id over various types)\n\n\n\n\n\n\nmy segmentation net learns local waveform and not the row label of lead signals ... i need to redesign my network ...", "votes": 1 }, { "id": 3312502, "postDate": "2025-11-07T10:45:12.257Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F353909259af828c279fb40275794aa76%2FSelection_922.png?generation=1762512097101606&alt=media\")
i realise that the main difficulty of applying supervised learning is creating the ground truth. There is always this one pixel shift error …
\ninstead of : model = learn (data, truth)
\nwe have: model = learn (data, not perfect truth)
Hence the learning is really, repeat:
\nmodel, better truth = learn (data, not perfect truth)
learning the ground truth is something unique about this competition
", "rawMarkdown": "\n\ni realise that the main difficulty of applying supervised learning is creating the ground truth. There is always this one pixel shift error ... \ninstead of : model = learn (data, truth) \nwe have: model = learn (data, not perfect truth)\n\n\nHence the learning is really, repeat: \nmodel, better truth = learn (data, not perfect truth)\n\nlearning the ground truth is something unique about this competition\n", "votes": 1 }, { "id": 3311453, "postDate": "2025-11-05T03:42:05.467Z", "content": "easiest way to extract ground truth mask for training, use read channel.\nmarker image made by averging red channel. \nnote the time … that is interesting
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe63c7b7352e5156537b2955ba2aa9c07%2FSelection_901.png?generation=1762314112584515&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fdcc53af015af148507abbdb82db6de3c%2FSelection_902.png?generation=1762314177783996&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fde8c24785d2852a09fa66ab611a08286%2F4247903125.png?generation=1762314316574102&alt=media\")
Which ECG option did you use, the one that 0001?
", "rawMarkdown": "Which ECG option did you use, the one that 0001?", "replies": [ { "id": 3312685, "postDate": "2025-11-07T17:46:35.600Z", "content": "yes. alternatively do the following:\n1) read series from csv file\n2) make an empty plot image (np array) of size = height x length of signal\n3) use opencv to draw 1 pixel thickness line with antialising connect the series \n4) resize to ecg image using INTER_AREA
\nthis is useful when there is overlap of different leads and doesn't over lap with the words, etc
", "rawMarkdown": "yes. alternatively do the following:\n1) read series from csv file\n2) make an empty plot image (np array) of size = height x length of signal\n3) use opencv to draw 1 pixel thickness line with antialising connect the series \n4) resize to ecg image using INTER_AREA\n\nthis is useful when there is overlap of different leads and doesn't over lap with the words, etc", "votes": 2 } ] } ] }, { "id": 3310046, "postDate": "2025-11-01T23:48:21.247Z", "content": "post processing …\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F7fa5c70394053f67a200b5ab807b0aec%2FSelection_872.png?generation=1762040866614935&alt=media\")
\nand i learn a new trick:\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fadfb363e1fe131c7013a90b985f29de6%2FSelection_871.png?generation=1762040878452488&alt=media\")
i wonder will this work:
\n1) given: image I, gridpoint_xy
\n2) create referennce R0. Then create R1 R2… are subpixel (or pixel) shifted versions of R0
\n3) rectified I to each of the R to get M0,M1,M2 …
\n4) predict series …. S0,S1,S2 … from unet(M0),unet(M1), … or unet(M0,M1,M2)?
\n5) a combination net: final = combine( S0,S1,S2 …)
in a perfect setting, the shift in S0,S1,S2 will be same as R0, R1 R2…. because there are error in rectification, mask prediction, etc … maybe we can learned the pattern of error and correct it?
", "rawMarkdown": "i wonder will this work: \n1) given: image I, gridpoint_xy \n2) create referennce R0. Then create R1 R2... are subpixel (or pixel) shifted versions of R0 \n3) rectified I to each of the R to get M0,M1,M2 ... \n4) predict series .... S0,S1,S2 ... from unet(M0),unet(M1), ... or unet(M0,M1,M2)? \n5) a combination net: final = combine( S0,S1,S2 ...) \n\nin a perfect setting, the shift in S0,S1,S2 will be same as R0, R1 R2.... because there are error in rectification, mask prediction, etc ... maybe we can learned the pattern of error and correct it?", "votes": 2 }, { "id": 3308108, "postDate": "2025-10-28T15:45:40.877Z", "content": "the takeaway message is that if the lead has high frequency signal (sharp peak), maybe you can stretch the input for better snr?
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe4e6e0a9e5decd9765c2829669777159%2FSelection_844.png?generation=1761666238433569&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F845558c683f85bbb04b59de05aa0455a%2FSelection_843.png?generation=1761666255439826&alt=media\")
Interesting. Yeah, the one thing I saw from the winning solution in the 2024 challenge was the sharp peaks being smoothed a bit. \nThis seems like a nice way around it!
", "rawMarkdown": "Interesting. Yeah, the one thing I saw from the winning solution in the 2024 challenge was the sharp peaks being smoothed a bit. \nThis seems like a nice way around it!", "votes": 1 } ] }, { "id": 3308090, "postDate": "2025-10-28T14:44:41.827Z", "content": "there is actually a shortcut. we can skip all unwarping of distorted grpah paper.
\nsince we have infinite train data, we predict the mv and sec values for each pixel directly.
Hence just a unet: image --> [Unet] --> for each pixel, lead signal label, mv and sec values.
\nThink of it as UVmap prediction for texturing in deep nets for computer graphics.
\nBut this require many many samples.
How can i build this UNET? can u please guide me through this.\nI referred to this paper https://arxiv.org/pdf/2510.19590 (opens in a new tab)\">https://arxiv.org/pdf/2510.19590 \nI created the UNET according to the weights mentioned in the paper.\nI gave input of 1 png to the UNET after transforming it using \"torchvision.transforms\"\nand the masked output is then reverse transformed and checked.
\nThe mask image is completely black.
\nwhere am I going wrong how can I solve this.
", "rawMarkdown": "How can i build this UNET? can u please guide me through this.\nI referred to this paper [https://arxiv.org/pdf/2510.19590 ](url)\nI created the UNET according to the weights mentioned in the paper.\nI gave input of 1 png to the UNET after transforming it using \"torchvision.transforms\"\nand the masked output is then reverse transformed and checked.\n\nThe mask image is completely black.\n\nwhere am I going wrong how can I solve this.", "replies": [ { "id": 3348374, "postDate": "2025-11-25T19:28:01.613Z", "content": "I would suggest taking a look here to see how the UNet is used: https://www.kaggle.com/code/eliasstenhede/open-ecg-digitizer
\nIt is the same network and weights as in the arXiv paper you refer to.
", "rawMarkdown": "I would suggest taking a look here to see how the UNet is used: https://www.kaggle.com/code/eliasstenhede/open-ecg-digitizer\n\nIt is the same network and weights as in the arXiv paper you refer to.", "votes": 1 } ] } ] }, { "id": 3311137, "postDate": "2025-11-04T10:37:06.917Z", "content": "winning trick:\nII signal appears twice\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F8a472009d49a1c781789fda860f3f824%2FSelection_899.png?generation=1762252624924541&alt=media\")
Do you train the segmentation model using a self-created dataset or a public dataset? If you use a public dataset, please tell me which one it is.
", "rawMarkdown": "Do you train the segmentation model using a self-created dataset or a public dataset? If you use a public dataset, please tell me which one it is.", "replies": [ { "id": 3311214, "postDate": "2025-11-04T13:52:19.500Z", "content": "\"Does the *.gridpoint_xy.npy file store the coordinates of grid points detected by Unet, or something else?\"
", "rawMarkdown": "\"Does the *.gridpoint_xy.npy file store the coordinates of grid points detected by Unet, or something else?\"" } ] } ] }, { "id": 3311081, "postDate": "2025-11-04T07:48:41.693Z", "content": "perlin noise by chatgpt
\n!pip install \nfrom import pnoise2\n\ndef perlin_noise_2d(\n H, W,\n =None\n):\n = np.random.(,) #\n octaves= np.random.randint(,) # #detail of \n persistence=np.random.(,) #,\n lacunarity=np.random.(,) #,\n\n is None:\n = np.random.randint(, )\n\n arr = np.zeros((H, W), dtype=np.float32)\n i range(H):\n j range(W):\n arr[i, j] = pnoise2(\n i / , j / ,\n octaves=octaves,\n persistence=persistence,\n lacunarity=lacunarity,\n repeatx=, repeaty=,\n base=,\n )\n # Normalize to [,]\n arr = (arr - arr.()) / (arr.() - arr.())\n arr\n\ndef make_dirt_patch():\n = perlin_noise_2d(, )\n\n # random nonlinear contrast\n = np.power(, np.random.(, ))\n # random threshold to isolate blobs\n thresh = np.random.(, )\n mask = ( > thresh).astype(np.float32)\n\n # smooth edges\n k = np.random.randint(, )\n mask = cv2.GaussianBlur(mask, (k | , k | ), )\n dirt = - mask * ()\n dirt\n\n# usage -----------------------------------------------------------------\n :\n = \n patch = make_dirt_patch()\n x = (CS-)\n y = (CS-)\n mask = np.ones((CS,CS,), np.float32)\n mask[y:y + , x:x + ] = patch[...,np.newaxis]\n augmented = *mask\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F3d7eaee2c999adc0abed2de8ea32d814%2FSelection_894.png?generation=1762242574869149&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F2e3f48fad4c12f31bc960c8e0ec3b815%2FSelection_895.png?generation=1762242598433623&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ffec95c9366bf835c3ec8b5ae860f9a34%2FSelection_893.png?generation=1762242612994223&alt=media\")
Thank you for sharing your great solution. Could you please explain what the 8 orientation classes in the stage0 model represent?
", "rawMarkdown": "Thank you for sharing your great solution. Could you please explain what the 8 orientation classes in the stage0 model represent?" }, { "id": 3379297, "postDate": "2025-12-19T14:42:30.293Z", "content": "uh,Sorry, I'm a bit curious, where does the ground truth of the line or grid come from?
", "rawMarkdown": "uh,Sorry, I'm a bit curious, where does the ground truth of the line or grid come from?" }, { "id": 3311056, "postDate": "2025-11-04T05:58:34.197Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe0e4befbd58600f7b27f8b232db97442%2FSelection_897.png?generation=1762246323407254&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F35759caca3cb1ef1b7aa592735e3553b%2FSelection_896.png?generation=1762246336532359&alt=media\")
\ni am quite impressed … need an iterative head (maybe a PPDM like head)
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fec4e122a7aa27a19ddb39be8b7b05640%2FSelection_891.png?generation=1762237028098545&alt=media\")
one ot two years later, chatgpt will code for me and show me the results
", "rawMarkdown": "\n\none ot two years later, chatgpt will code for me and show me the results", "votes": 2 } ] }, { "id": 3309777, "postDate": "2025-11-01T11:28:36.733Z", "content": "Sorry, but where can I see the last year's winning solution…?
", "rawMarkdown": "Sorry, but where can I see the last year's winning solution...?" }, { "id": 3309541, "postDate": "2025-11-01T00:19:16.753Z", "content": "maybe useful:\n1) https://github.com/zwxu064/DiffImageWarpingCUDA\ndifferentiable pixel warping
", "rawMarkdown": "maybe useful:\n1) https://github.com/zwxu064/DiffImageWarpingCUDA\ndifferentiable pixel warping\n" }, { "id": 3309512, "postDate": "2025-10-31T22:25:04.157Z", "content": "I think the SNR loss is a good approach. You might find this unet+weights useful for further fine-tuning:
\nhttps://github.com/Ahus-AIM/Open-ECG-Digitizer/blob/main/src/model/unet.py
\nIt segments images into four classes: background, text, ECG signal and grid lines.
", "rawMarkdown": "I think the SNR loss is a good approach. You might find this unet+weights useful for further fine-tuning:\n\nhttps://github.com/Ahus-AIM/Open-ECG-Digitizer/blob/main/src/model/unet.py\n\nIt segments images into four classes: background, text, ECG signal and grid lines." }, { "id": 3308601, "postDate": "2025-10-29T18:16:08.637Z", "content": "What if to plot time series to generate ground truth images. \nThen train image 2 image translation from raw images to this ground truth image and have some AUX regression loss from the translated image to the time series data.
", "rawMarkdown": "What if to plot time series to generate ground truth images. \nThen train image 2 image translation from raw images to this ground truth image and have some AUX regression loss from the translated image to the time series data." }, { "id": 3308267, "postDate": "2025-10-29T01:56:20.210Z", "content": "Is it possible for other projects as well? like Brain Stroke @hengck23
", "rawMarkdown": "Is it possible for other projects as well? like Brain Stroke @hengck23 " }, { "id": 3308252, "postDate": "2025-10-29T00:36:39.933Z", "content": "Hi,
\nJust out of curiosity, is the idea mainly trying to segment the signals from the images?
", "rawMarkdown": "Hi,\n\nJust out of curiosity, is the idea mainly trying to segment the signals from the images?", "replies": [ { "id": 3308256, "postDate": "2025-10-29T00:56:09.913Z", "content": "i think the main difficulty of the challenge is how to warp the image correctly. wrong wrapping affect t the mV values and hence low snr.
", "rawMarkdown": "i think the main difficulty of the challenge is how to warp the image correctly. wrong wrapping affect t the mV values and hence low snr.", "votes": 2, "replies": [ { "id": 3308887, "postDate": "2025-10-30T13:40:06.203Z", "content": "It seems so.
\nI took a look on the dataset and wraping different images regardless of perspective or introduced artifacts is challenging.\nI recall you shared a post about a paper doing exactly this using a Gridder Algorithm. It is mainly trained on a small sample of ECG images.\nPerhaps training a keypoint detection model like YOLO on some annotated samples of ECG images (keypoints are on the intersections of the 5mm x 5mm grid), it may increase the SNR performance.
\nI'm curious about your thoughts on this!
", "rawMarkdown": "It seems so. \n\nI took a look on the dataset and wraping different images regardless of perspective or introduced artifacts is challenging.\nI recall you shared a post about a paper doing exactly this using a **Gridder Algorithm**. It is mainly trained on a small sample of ECG images.\nPerhaps training a keypoint detection model like YOLO on some annotated samples of ECG images (keypoints are on the intersections of the 5mm x 5mm grid), it may increase the SNR performance.\n\nI'm curious about your thoughts on this!", "votes": 1 }, { "id": 3308967, "postDate": "2025-10-30T16:08:17.577Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4746c8d75efd05b447fc3de14bac6b02%2FSelection_854.png?generation=1761840244179204&alt=media\")
for me the flow is :
\nVery simple approach! 👍
", "rawMarkdown": "Very simple approach! 👍" } ] } ] }, { "id": 3307865, "postDate": "2025-10-28T02:00:00.083Z", "content": "best pipeline i can think of now? This is not confirmed … i am choosing between labeling the grid lines (line1, line2 …. line50) on the whole image or the local crop.
\n1 ) input image --> label pixel for
\n2) crop each signal
\nhost tries to make my life difficult 😭
\n
😂 oh multi pages, maybe we need to get the keypoints for foreground page first. By the way, are the green points produced by your segmentation model?
", "votes": 0, "replies": [ { "id": 3310643, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-03T10:25:53.807000", "content": "all results are raw results from the segmentation model without post processing yet. i am trying to simplify the steps, i can add a foreground object detection … but this add another step.
", "votes": 0, "replies": [] }, { "id": 3318741, "author_name": "", "author_url": "", "post_date": "2025-11-11T10:55:08.587000", "content": "", "votes": 0, "replies": [] }, { "id": 3318742, "author_name": "Darshan Makwana", "author_url": "", "post_date": "2025-11-11T10:55:36.927000", "content": "I see the foreground grid points are not connected to the main grid points, wouldn't removing the connected component with the least grid points work in your case?
", "votes": 0, "replies": [] } ] } ] }, { "id": 3308257, "author_name": "hengck23", "author_url": "", "post_date": "2025-10-29T00:57:11.983000", "content": "my unwarping experiment is a success ?\npreliminary results, code to come …
\n
2. Tricks and Pitfalls
\n[truth label]
\n
[predict label]
\n[context window]
\n
[missing prediction]
\n[dataset infor]
\n\"Can we expect the public/private test to have a uniform distribution of the 9 augmentation types? Or is the test assembled using only type 1 like the two examples? Or was the test even constructed using augmented samples not present in the training set, such as types 2, 7, 8?\"
\nresults on validation (tarined with 100 image id over various types)
\n\n
my segmentation net learns local waveform and not the row label of lead signals … i need to redesign my network …
", "votes": 1, "replies": [] }, { "id": 3312502, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-07T10:45:12.257000", "content": "
i realise that the main difficulty of applying supervised learning is creating the ground truth. There is always this one pixel shift error …
\ninstead of : model = learn (data, truth)
\nwe have: model = learn (data, not perfect truth)
Hence the learning is really, repeat:
\nmodel, better truth = learn (data, not perfect truth)
learning the ground truth is something unique about this competition
", "votes": 1, "replies": [] }, { "id": 3311453, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-05T03:42:05.467000", "content": "easiest way to extract ground truth mask for training, use read channel.\nmarker image made by averging red channel. \nnote the time … that is interesting
\n
Which ECG option did you use, the one that 0001?
", "votes": 0, "replies": [ { "id": 3312685, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-07T17:46:35.600000", "content": "yes. alternatively do the following:\n1) read series from csv file\n2) make an empty plot image (np array) of size = height x length of signal\n3) use opencv to draw 1 pixel thickness line with antialising connect the series \n4) resize to ecg image using INTER_AREA
\nthis is useful when there is overlap of different leads and doesn't over lap with the words, etc
", "votes": 2, "replies": [] } ] } ] }, { "id": 3310046, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-01T23:48:21.247000", "content": "post processing …\n\nand i learn a new trick:\n
i wonder will this work:
\n1) given: image I, gridpoint_xy
\n2) create referennce R0. Then create R1 R2… are subpixel (or pixel) shifted versions of R0
\n3) rectified I to each of the R to get M0,M1,M2 …
\n4) predict series …. S0,S1,S2 … from unet(M0),unet(M1), … or unet(M0,M1,M2)?
\n5) a combination net: final = combine( S0,S1,S2 …)
in a perfect setting, the shift in S0,S1,S2 will be same as R0, R1 R2…. because there are error in rectification, mask prediction, etc … maybe we can learned the pattern of error and correct it?
", "votes": 2, "replies": [] }, { "id": 3308108, "author_name": "hengck23", "author_url": "", "post_date": "2025-10-28T15:45:40.877000", "content": "the takeaway message is that if the lead has high frequency signal (sharp peak), maybe you can stretch the input for better snr?
\n\n
Interesting. Yeah, the one thing I saw from the winning solution in the 2024 challenge was the sharp peaks being smoothed a bit. \nThis seems like a nice way around it!
", "votes": 1, "replies": [] } ] }, { "id": 3308090, "author_name": "hengck23", "author_url": "", "post_date": "2025-10-28T14:44:41.827000", "content": "there is actually a shortcut. we can skip all unwarping of distorted grpah paper.
\nsince we have infinite train data, we predict the mv and sec values for each pixel directly.
Hence just a unet: image --> [Unet] --> for each pixel, lead signal label, mv and sec values.
\nThink of it as UVmap prediction for texturing in deep nets for computer graphics.
\nBut this require many many samples.
How can i build this UNET? can u please guide me through this.\nI referred to this paper https://arxiv.org/pdf/2510.19590 (opens in a new tab)\">https://arxiv.org/pdf/2510.19590 \nI created the UNET according to the weights mentioned in the paper.\nI gave input of 1 png to the UNET after transforming it using \"torchvision.transforms\"\nand the masked output is then reverse transformed and checked.
\nThe mask image is completely black.
\nwhere am I going wrong how can I solve this.
", "votes": 0, "replies": [ { "id": 3348374, "author_name": "Elias Stenhede", "author_url": "", "post_date": "2025-11-25T19:28:01.613000", "content": "I would suggest taking a look here to see how the UNet is used: https://www.kaggle.com/code/eliasstenhede/open-ecg-digitizer
\nIt is the same network and weights as in the arXiv paper you refer to.
", "votes": 1, "replies": [] } ] } ] }, { "id": 3311137, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-04T10:37:06.917000", "content": "winning trick:\nII signal appears twice\n
Do you train the segmentation model using a self-created dataset or a public dataset? If you use a public dataset, please tell me which one it is.
", "votes": 0, "replies": [ { "id": 3311214, "author_name": "shixingkong", "author_url": "", "post_date": "2025-11-04T13:52:19.500000", "content": "\"Does the *.gridpoint_xy.npy file store the coordinates of grid points detected by Unet, or something else?\"
", "votes": 0, "replies": [] } ] } ] }, { "id": 3311081, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-04T07:48:41.693000", "content": "perlin noise by chatgpt
\n!pip install \nfrom import pnoise2\n\ndef perlin_noise_2d(\n H, W,\n =None\n):\n = np.random.(,) #\n octaves= np.random.randint(,) # #detail of \n persistence=np.random.(,) #,\n lacunarity=np.random.(,) #,\n\n is None:\n = np.random.randint(, )\n\n arr = np.zeros((H, W), dtype=np.float32)\n i range(H):\n j range(W):\n arr[i, j] = pnoise2(\n i / , j / ,\n octaves=octaves,\n persistence=persistence,\n lacunarity=lacunarity,\n repeatx=, repeaty=,\n base=,\n )\n # Normalize to [,]\n arr = (arr - arr.()) / (arr.() - arr.())\n arr\n\ndef make_dirt_patch():\n = perlin_noise_2d(, )\n\n # random nonlinear contrast\n = np.power(, np.random.(, ))\n # random threshold to isolate blobs\n thresh = np.random.(, )\n mask = ( > thresh).astype(np.float32)\n\n # smooth edges\n k = np.random.randint(, )\n mask = cv2.GaussianBlur(mask, (k | , k | ), )\n dirt = - mask * ()\n dirt\n\n# usage -----------------------------------------------------------------\n :\n = \n patch = make_dirt_patch()\n x = (CS-)\n y = (CS-)\n mask = np.ones((CS,CS,), np.float32)\n mask[y:y + , x:x + ] = patch[...,np.newaxis]\n augmented = *mask\n
Thank you for sharing your great solution. Could you please explain what the 8 orientation classes in the stage0 model represent?
", "votes": 0, "replies": [] }, { "id": 3379297, "author_name": "yongchen918", "author_url": "", "post_date": "2025-12-19T14:42:30.293000", "content": "uh,Sorry, I'm a bit curious, where does the ground truth of the line or grid come from?
", "votes": 0, "replies": [] }, { "id": 3311056, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-04T05:58:34.197000", "content": "\n\ni am quite impressed … need an iterative head (maybe a PPDM like head)
one ot two years later, chatgpt will code for me and show me the results
", "votes": 2, "replies": [] } ] }, { "id": 3309777, "author_name": "DJ Moon", "author_url": "", "post_date": "2025-11-01T11:28:36.733000", "content": "Sorry, but where can I see the last year's winning solution…?
", "votes": 0, "replies": [] }, { "id": 3309541, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-01T00:19:16.753000", "content": "maybe useful:\n1) https://github.com/zwxu064/DiffImageWarpingCUDA\ndifferentiable pixel warping
", "votes": 0, "replies": [] }, { "id": 3309512, "author_name": "Elias Stenhede", "author_url": "", "post_date": "2025-10-31T22:25:04.157000", "content": "I think the SNR loss is a good approach. You might find this unet+weights useful for further fine-tuning:
\nhttps://github.com/Ahus-AIM/Open-ECG-Digitizer/blob/main/src/model/unet.py
\nIt segments images into four classes: background, text, ECG signal and grid lines.
", "votes": 0, "replies": [] }, { "id": 3308601, "author_name": "Maxim Shatskiy", "author_url": "", "post_date": "2025-10-29T18:16:08.637000", "content": "What if to plot time series to generate ground truth images. \nThen train image 2 image translation from raw images to this ground truth image and have some AUX regression loss from the translated image to the time series data.
", "votes": 0, "replies": [] }, { "id": 3308267, "author_name": "Navneet", "author_url": "", "post_date": "2025-10-29T01:56:20.210000", "content": "Is it possible for other projects as well? like Brain Stroke @hengck23
", "votes": 0, "replies": [] }, { "id": 3308252, "author_name": "DJ Moon", "author_url": "", "post_date": "2025-10-29T00:36:39.933000", "content": "Hi,
\nJust out of curiosity, is the idea mainly trying to segment the signals from the images?
", "votes": 0, "replies": [ { "id": 3308256, "author_name": "hengck23", "author_url": "", "post_date": "2025-10-29T00:56:09.913000", "content": "i think the main difficulty of the challenge is how to warp the image correctly. wrong wrapping affect t the mV values and hence low snr.
", "votes": 2, "replies": [ { "id": 3308887, "author_name": "Marouane El Massine Lfellah", "author_url": "", "post_date": "2025-10-30T13:40:06.203000", "content": "It seems so.
\nI took a look on the dataset and wraping different images regardless of perspective or introduced artifacts is challenging.\nI recall you shared a post about a paper doing exactly this using a Gridder Algorithm. It is mainly trained on a small sample of ECG images.\nPerhaps training a keypoint detection model like YOLO on some annotated samples of ECG images (keypoints are on the intersections of the 5mm x 5mm grid), it may increase the SNR performance.
\nI'm curious about your thoughts on this!
", "votes": 1, "replies": [] }, { "id": 3308967, "author_name": "hengck23", "author_url": "", "post_date": "2025-10-30T16:08:17.577000", "content": "
for me the flow is :
\nVery simple approach! 👍
", "votes": 0, "replies": [] } ] } ] }, { "id": 3307865, "author_name": "hengck23", "author_url": "", "post_date": "2025-10-28T02:00:00.083000", "content": "best pipeline i can think of now? This is not confirmed … i am choosing between labeling the grid lines (line1, line2 …. line50) on the whole image or the local crop.
\n1 ) input image --> label pixel for
\n2) crop each signal
\n