{ "id": 302586, "title": "SAHI - new tutorial and notebook - Slicing Aided Hyper Inference for Yolov5 and YoloX", "url": "/competitions/tensorflow-great-barrier-reef/discussion/302586", "author_name": "Remek Kinas", "post_date": "2022-01-23T08:12:47.900000", "votes": 68, "comment_count": 36, "views": 0, "content": "

Today I provided new notebook for experimentations - SAHI - Slicing Aided Hyper Inference. SAHI is a vision library for large-scale object detection & instance segmentation. Now I introduced it on Kaggle for Yolov5 (YoloX will be soon - testing it). I think it could be helful for us - we deal with many small object here ….

\n

Hope you enjoy it and find it helpful not only in this comeptition but in your work/other competitions.

\n

Notebook: https://www.kaggle.com/remekkinas/sahi-slicing-aided-hyper-inference-yv5-and-yx

\n

\"video\"

\n

\"SAHI\"

", "messages": [ { "id": 1661094, "postDate": "2022-01-23T08:12:47.900Z", "content": "

Today I provided new notebook for experimentations - SAHI - Slicing Aided Hyper Inference. SAHI is a vision library for large-scale object detection & instance segmentation. Now I introduced it on Kaggle for Yolov5 (YoloX will be soon - testing it). I think it could be helful for us - we deal with many small object here ….

\n

Hope you enjoy it and find it helpful not only in this comeptition but in your work/other competitions.

\n

Notebook: https://www.kaggle.com/remekkinas/sahi-slicing-aided-hyper-inference-yv5-and-yx

\n

\"video\"

\n

\"SAHI\"

", "rawMarkdown": "Today I provided new notebook for experimentations - **SAHI - Slicing Aided Hyper Inference. SAHI is a vision library for large-scale object detection & instance segmentation**. Now I introduced it on Kaggle for Yolov5 (YoloX will be soon - testing it). I think it could be helful for us - we deal with many small object here ....\n\nHope you enjoy it and find it helpful not only in this comeptition but in your work/other competitions.\n\nNotebook: https://www.kaggle.com/remekkinas/sahi-slicing-aided-hyper-inference-yv5-and-yx\n\n\n![video](https://youtu.be/gUex8YoguHU)\n\n\n\n![SAHI](https://user-images.githubusercontent.com/34196005/144092739-c1d9bade-a128-4346-947f-424ce00e5c4f.gif)", "votes": 68 }, { "id": 1663537, "postDate": "2022-01-25T07:34:33.783Z", "content": "

Did sahi really work?👀

", "rawMarkdown": "Did sahi really work?👀", "votes": 1, "replies": [ { "id": 1663555, "postDate": "2022-01-25T07:58:06.590Z", "content": "

For me it didn't. Lowered LB from 0.547 to 0.496 for one model. But it is a wonderful idea and would be a great tool to use for training AND inference

", "rawMarkdown": "For me it didn't. Lowered LB from 0.547 to 0.496 for one model. But it is a wonderful idea and would be a great tool to use for training AND inference" } ] }, { "id": 1662218, "postDate": "2022-01-24T06:00:45.247Z", "content": "

Thank You For This. Actually i have started learning ML Algorithms from past 15-20 days. But my concepts about what happens in the backend are not getting clear. So if you can guide me on this then it will be very helpful. Or the way you have done ML Algo's Deeply. Basically study material. Once again Thanks alot.

", "rawMarkdown": "Thank You For This. Actually i have started learning ML Algorithms from past 15-20 days. But my concepts about what happens in the backend are not getting clear. So if you can guide me on this then it will be very helpful. Or the way you have done ML Algo's Deeply. Basically study material. Once again Thanks alot.", "votes": 1 }, { "id": 1662183, "postDate": "2022-01-24T05:19:41.977Z", "content": "

Great ! I am not satisfied with SAHI because it never improved my score

", "rawMarkdown": "Great ! I am not satisfied with SAHI because it never improved my score\n", "votes": 1, "replies": [ { "id": 1662533, "postDate": "2022-01-24T11:12:07.280Z", "content": "

Key is to understad how SAHI works (as all tools we sue). I play many times with SAHI with success … this case however (bad prediction) forced me to understand what is under hood … and … I think it will finally work :) I am … implementing custom solution which will be availabe soon.

", "rawMarkdown": "Key is to understad how SAHI works (as all tools we sue). I play many times with SAHI with success ... this case however (bad prediction) forced me to understand what is under hood ... and ... I think it will finally work :) I am ... implementing custom solution which will be availabe soon. ", "votes": 2 }, { "id": 1662928, "postDate": "2022-01-24T16:49:25.403Z", "content": "

I didn't experiment a lot but saw it don't from initial experiments. i dropped that idea. I implemented my own tiling algorithm.

", "rawMarkdown": "I didn't experiment a lot but saw it don't from initial experiments. i dropped that idea. I implemented my own tiling algorithm." } ] }, { "id": 1661777, "postDate": "2022-01-23T17:33:34.987Z", "content": "

@remekkinas Beat me to a Sahi notebook, I am waiting for a good score on my model ;). Great job!

", "rawMarkdown": "@remekkinas Beat me to a Sahi notebook, I am waiting for a good score on my model ;). Great job!", "votes": 1 }, { "id": 1661501, "postDate": "2022-01-23T13:56:54.170Z", "content": "

Thanks for sharing. Waiting for yolox!

", "rawMarkdown": "Thanks for sharing. Waiting for yolox!", "votes": 1, "replies": [ { "id": 1661540, "postDate": "2022-01-23T14:14:26.843Z", "content": "

Working on it … :)

", "rawMarkdown": "Working on it ... :)" } ] }, { "id": 1661423, "postDate": "2022-01-23T13:12:47.607Z", "content": "

Thank you for this notebook! And thank you for all your nice sharings during the competition!

", "rawMarkdown": "Thank you for this notebook! And thank you for all your nice sharings during the competition!", "votes": 1, "replies": [ { "id": 1661429, "postDate": "2022-01-23T13:14:39.910Z", "content": "

I am so happy that this is valuable for you. Thank you for nice words. This motivate me to share.

", "rawMarkdown": "I am so happy that this is valuable for you. Thank you for nice words. This motivate me to share.", "votes": 1 } ] }, { "id": 1661195, "postDate": "2022-01-23T09:56:56.370Z", "content": "

Thanks for all the amazing notebooks you have been creating for the competition. I've learned a lot from you. @remekkinas ❤️

", "rawMarkdown": "Thanks for all the amazing notebooks you have been creating for the competition. I've learned a lot from you. @remekkinas ❤️", "votes": 1, "replies": [ { "id": 1661203, "postDate": "2022-01-23T10:01:41.387Z", "content": "

Thank you very much. It is nice! ❤️😍

", "rawMarkdown": "Thank you very much. It is nice! ❤️😍", "votes": 1 } ] }, { "id": 1661191, "postDate": "2022-01-23T09:52:22.193Z", "content": "

I was just trying this on collab… thnx anyways

", "rawMarkdown": "I was just trying this on collab... thnx anyways", "votes": 1, "replies": [ { "id": 1661200, "postDate": "2022-01-23T10:00:50.937Z", "content": "

Cool! It is really great tool we can use not only in this completion this is why I decided to share.

", "rawMarkdown": "Cool! It is really great tool we can use not only in this completion this is why I decided to share." }, { "id": 1661261, "postDate": "2022-01-23T10:38:38.593Z", "content": "

Yes, I was doing it manually and creating 4 sections and modifying labels accordingly it was such a painful task… Hope you shared it earlier it would have saved me a lot of hours 🥲. Let me know if you have any other tricks up your sleeves 😂

", "rawMarkdown": "Yes, I was doing it manually and creating 4 sections and modifying labels accordingly it was such a painful task... Hope you shared it earlier it would have saved me a lot of hours 🥲. Let me know if you have any other tricks up your sleeves 😂", "votes": 1 }, { "id": 1661481, "postDate": "2022-01-23T13:42:28.997Z", "content": "

https://github.com/obss/sahi/blob/main/demo/slicing.ipynb
\nYou can easily slice training images with this code and set whether save negative examples. Also, sliced training brought me more FPs, I haven't figured out how to deal with it.

", "rawMarkdown": "https://github.com/obss/sahi/blob/main/demo/slicing.ipynb\nYou can easily slice training images with this code and set whether save negative examples. Also, sliced training brought me more FPs, I haven't figured out how to deal with it.", "votes": 1 }, { "id": 1662672, "postDate": "2022-01-24T13:09:20.900Z", "content": "

@zzy990106 look in new release I provided minutes ago. I think it gave you more possibilities to experiment. Let me know if I can do something better or improve notebook. Thanks!

", "rawMarkdown": "@zzy990106 look in new release I provided minutes ago. I think it gave you more possibilities to experiment. Let me know if I can do something better or improve notebook. Thanks!", "votes": 1 } ] }, { "id": 1661155, "postDate": "2022-01-23T09:11:44.433Z", "content": "

very nice!

", "rawMarkdown": "very nice!", "votes": 1, "replies": [ { "id": 1661159, "postDate": "2022-01-23T09:15:43.317Z", "content": "

Thank you! :) I hope you will benefit from this tool 👍

", "rawMarkdown": "Thank you! :) I hope you will benefit from this tool 👍" } ] }, { "id": 1662597, "postDate": "2022-01-24T12:14:22.693Z", "content": "

That's really interesting. Amazing work overall @remekkinas

", "rawMarkdown": "That's really interesting. Amazing work overall @remekkinas ", "replies": [ { "id": 1662668, "postDate": "2022-01-24T13:07:54.560Z", "content": "

@crained now I released improved notebook - custom prediction class. It is more flexible and better to perform sophisticated experimentations. Way better :)

\n

I am sure I found way to improve score …. implementing now.

", "rawMarkdown": "@crained now I released improved notebook - custom prediction class. It is more flexible and better to perform sophisticated experimentations. Way better :) \n\nI am sure I found way to improve score .... implementing now. ", "votes": 1 }, { "id": 1662676, "postDate": "2022-01-24T13:11:23.917Z", "content": "

That's great. Especially considering your amazing position currently 😊

", "rawMarkdown": "That's great. Especially considering your amazing position currently 😊" }, { "id": 1662684, "postDate": "2022-01-24T13:17:58.160Z", "content": "

Do not have many GPUs to train model I dream about :)

", "rawMarkdown": "Do not have many GPUs to train model I dream about :) " } ] }, { "id": 1694946, "postDate": "2022-02-17T21:04:25.787Z", "content": "

As far as I can see now SAHI paper is available here: https://arxiv.org/abs/2202.06934

", "rawMarkdown": "As far as I can see now SAHI paper is available here: https://arxiv.org/abs/2202.06934" }, { "id": 1694668, "postDate": "2022-02-17T16:13:27.723Z", "content": "

@remekkinas what are your observations after the competition,
\nI implemented sliced inference (SAHI) during the competition but it was performing poorly on public leaderboard and thus didn't experiment much on it,however in observations after the competition it gave some large marginal jump in score from public to private leaderboard ( as high as + 0.16 )
\nI guess this algorithm has great potential in object detection just needs more research :)

", "rawMarkdown": "@remekkinas what are your observations after the competition,\nI implemented sliced inference (SAHI) during the competition but it was performing poorly on public leaderboard and thus didn't experiment much on it,however in observations after the competition it gave some large marginal jump in score from public to private leaderboard ( as high as + 0.16 )\nI guess this algorithm has great potential in object detection just needs more research :)", "replies": [ { "id": 1694942, "postDate": "2022-02-17T21:02:28.097Z", "content": "

I spent too little time with SAHI here. Unfortunately I was really limited by my GPU … (we created all our solution on Colab/Notebook). I do not have time to discover it potential unfortunately :( My lessons learned for nex competition is to find machine with 2xGPU … (decrease training time, allow me to make more experiments - play with batch_size, other frameworks etc.).

", "rawMarkdown": "I spent too little time with SAHI here. Unfortunately I was really limited by my GPU ... (we created all our solution on Colab/Notebook). I do not have time to discover it potential unfortunately :( My lessons learned for nex competition is to find machine with 2xGPU ... (decrease training time, allow me to make more experiments - play with batch_size, other frameworks etc.)." } ] }, { "id": 1661466, "postDate": "2022-01-23T13:35:33.503Z", "rawMarkdown": "", "votes": 1, "isDeleted": true, "replies": [ { "id": 1661473, "postDate": "2022-01-23T13:38:41.500Z", "content": "

Idea is really good:

\n", "rawMarkdown": "Idea is really good:\n- take img - make a prediction on oryginal image (you can change disable this)\n- make a chips :) with width and height and predict on them \n- compute bboxes and scores for all predictions (this is the most complex part - because you make prediction on many chips - some objects could be on many chips at the same time) ", "votes": 1 }, { "id": 1661856, "postDate": "2022-01-23T18:52:36.820Z", "content": "

@remekkinas does the same slicing works with 1x images instead of 3x images in training?

", "rawMarkdown": "@remekkinas does the same slicing works with 1x images instead of 3x images in training?" }, { "id": 1661863, "postDate": "2022-01-23T19:04:16.857Z", "content": "

Training with different resolutions will affect the f2 score badly during prediction.

", "rawMarkdown": "Training with different resolutions will affect the f2 score badly during prediction. " }, { "id": 1661932, "postDate": "2022-01-23T20:33:00.527Z", "content": "

@anirudhyadav9784 you are right. I think we should train model differently for SAHI. I will test model creation this week. Idea behind this tool is great but it requires time (some experimentation) to achieve good score.

", "rawMarkdown": "@anirudhyadav9784 you are right. I think we should train model differently for SAHI. I will test model creation this week. Idea behind this tool is great but it requires time (some experimentation) to achieve good score." }, { "id": 1662283, "postDate": "2022-01-24T06:55:24.747Z", "content": "

Yep tried that but this is what I learned, Don't go beyond 4 slices even then false positive is the real pain. Need to train it much longer with optimized hyp for it to work else its cons will outrank its pros and seems like 30 hours Kaggle GPU limit is not enough for me.

", "rawMarkdown": "Yep tried that but this is what I learned, Don't go beyond 4 slices even then false positive is the real pain. Need to train it much longer with optimized hyp for it to work else its cons will outrank its pros and seems like 30 hours Kaggle GPU limit is not enough for me.", "votes": 1 } ] }, { "id": 1664087, "postDate": "2022-01-25T17:21:35.260Z", "content": "

Interesting work @remekkinas ✨Thanks!😄

", "rawMarkdown": "Interesting work @remekkinas ✨Thanks!😄", "votes": 1 }, { "id": 1662930, "postDate": "2022-01-24T16:49:34.870Z", "content": "

Thank You for sharing! Interesting!

", "rawMarkdown": "Thank You for sharing! Interesting!", "votes": 1 }, { "id": 1662881, "postDate": "2022-01-24T15:55:45.443Z", "content": "

Thank you for sharing!!! :D

", "rawMarkdown": "Thank you for sharing!!! :D", "votes": 1 } ], "comments": [ { "id": 1663537, "author_name": "Shichang Liu", "author_url": "", "post_date": "2022-01-25T07:34:33.783000", "content": "

Did sahi really work?👀

", "votes": 1, "replies": [ { "id": 1663555, "author_name": "nymfree", "author_url": "", "post_date": "2022-01-25T07:58:06.590000", "content": "

For me it didn't. Lowered LB from 0.547 to 0.496 for one model. But it is a wonderful idea and would be a great tool to use for training AND inference

", "votes": 0, "replies": [] } ] }, { "id": 1662218, "author_name": "lilzee", "author_url": "", "post_date": "2022-01-24T06:00:45.247000", "content": "

Thank You For This. Actually i have started learning ML Algorithms from past 15-20 days. But my concepts about what happens in the backend are not getting clear. So if you can guide me on this then it will be very helpful. Or the way you have done ML Algo's Deeply. Basically study material. Once again Thanks alot.

", "votes": 1, "replies": [] }, { "id": 1662183, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2022-01-24T05:19:41.977000", "content": "

Great ! I am not satisfied with SAHI because it never improved my score

", "votes": 1, "replies": [ { "id": 1662533, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-24T11:12:07.280000", "content": "

Key is to understad how SAHI works (as all tools we sue). I play many times with SAHI with success … this case however (bad prediction) forced me to understand what is under hood … and … I think it will finally work :) I am … implementing custom solution which will be availabe soon.

", "votes": 2, "replies": [] }, { "id": 1662928, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2022-01-24T16:49:25.403000", "content": "

I didn't experiment a lot but saw it don't from initial experiments. i dropped that idea. I implemented my own tiling algorithm.

", "votes": 0, "replies": [] } ] }, { "id": 1661777, "author_name": "outwrest", "author_url": "", "post_date": "2022-01-23T17:33:34.987000", "content": "

@remekkinas Beat me to a Sahi notebook, I am waiting for a good score on my model ;). Great job!

", "votes": 1, "replies": [] }, { "id": 1661501, "author_name": "Yangranran", "author_url": "", "post_date": "2022-01-23T13:56:54.170000", "content": "

Thanks for sharing. Waiting for yolox!

", "votes": 1, "replies": [ { "id": 1661540, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-23T14:14:26.843000", "content": "

Working on it … :)

", "votes": 0, "replies": [] } ] }, { "id": 1661423, "author_name": "Juntao Jiang", "author_url": "", "post_date": "2022-01-23T13:12:47.607000", "content": "

Thank you for this notebook! And thank you for all your nice sharings during the competition!

", "votes": 1, "replies": [ { "id": 1661429, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-23T13:14:39.910000", "content": "

I am so happy that this is valuable for you. Thank you for nice words. This motivate me to share.

", "votes": 1, "replies": [] } ] }, { "id": 1661195, "author_name": "Bao Loc Pham", "author_url": "", "post_date": "2022-01-23T09:56:56.370000", "content": "

Thanks for all the amazing notebooks you have been creating for the competition. I've learned a lot from you. @remekkinas ❤️

", "votes": 1, "replies": [ { "id": 1661203, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-23T10:01:41.387000", "content": "

Thank you very much. It is nice! ❤️😍

", "votes": 1, "replies": [] } ] }, { "id": 1661191, "author_name": "Anirudh", "author_url": "", "post_date": "2022-01-23T09:52:22.193000", "content": "

I was just trying this on collab… thnx anyways

", "votes": 1, "replies": [ { "id": 1661200, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-23T10:00:50.937000", "content": "

Cool! It is really great tool we can use not only in this completion this is why I decided to share.

", "votes": 0, "replies": [] }, { "id": 1661261, "author_name": "Anirudh", "author_url": "", "post_date": "2022-01-23T10:38:38.593000", "content": "

Yes, I was doing it manually and creating 4 sections and modifying labels accordingly it was such a painful task… Hope you shared it earlier it would have saved me a lot of hours 🥲. Let me know if you have any other tricks up your sleeves 😂

", "votes": 1, "replies": [] }, { "id": 1661481, "author_name": "Leon", "author_url": "", "post_date": "2022-01-23T13:42:28.997000", "content": "

https://github.com/obss/sahi/blob/main/demo/slicing.ipynb
\nYou can easily slice training images with this code and set whether save negative examples. Also, sliced training brought me more FPs, I haven't figured out how to deal with it.

", "votes": 1, "replies": [] }, { "id": 1662672, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-24T13:09:20.900000", "content": "

@zzy990106 look in new release I provided minutes ago. I think it gave you more possibilities to experiment. Let me know if I can do something better or improve notebook. Thanks!

", "votes": 1, "replies": [] } ] }, { "id": 1661155, "author_name": "tik_boa", "author_url": "", "post_date": "2022-01-23T09:11:44.433000", "content": "

very nice!

", "votes": 1, "replies": [ { "id": 1661159, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-23T09:15:43.317000", "content": "

Thank you! :) I hope you will benefit from this tool 👍

", "votes": 0, "replies": [] } ] }, { "id": 1662597, "author_name": "Charlie Craine", "author_url": "", "post_date": "2022-01-24T12:14:22.693000", "content": "

That's really interesting. Amazing work overall @remekkinas

", "votes": 0, "replies": [ { "id": 1662668, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-24T13:07:54.560000", "content": "

@crained now I released improved notebook - custom prediction class. It is more flexible and better to perform sophisticated experimentations. Way better :)

\n

I am sure I found way to improve score …. implementing now.

", "votes": 1, "replies": [] }, { "id": 1662676, "author_name": "Charlie Craine", "author_url": "", "post_date": "2022-01-24T13:11:23.917000", "content": "

That's great. Especially considering your amazing position currently 😊

", "votes": 0, "replies": [] }, { "id": 1662684, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-24T13:17:58.160000", "content": "

Do not have many GPUs to train model I dream about :)

", "votes": 0, "replies": [] } ] }, { "id": 1694946, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-02-17T21:04:25.787000", "content": "

As far as I can see now SAHI paper is available here: https://arxiv.org/abs/2202.06934

", "votes": 0, "replies": [] }, { "id": 1694668, "author_name": "th3y3llowbird", "author_url": "", "post_date": "2022-02-17T16:13:27.723000", "content": "

@remekkinas what are your observations after the competition,
\nI implemented sliced inference (SAHI) during the competition but it was performing poorly on public leaderboard and thus didn't experiment much on it,however in observations after the competition it gave some large marginal jump in score from public to private leaderboard ( as high as + 0.16 )
\nI guess this algorithm has great potential in object detection just needs more research :)

", "votes": 0, "replies": [ { "id": 1694942, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-02-17T21:02:28.097000", "content": "

I spent too little time with SAHI here. Unfortunately I was really limited by my GPU … (we created all our solution on Colab/Notebook). I do not have time to discover it potential unfortunately :( My lessons learned for nex competition is to find machine with 2xGPU … (decrease training time, allow me to make more experiments - play with batch_size, other frameworks etc.).

", "votes": 0, "replies": [] } ] }, { "id": 1661466, "author_name": "", "author_url": "", "post_date": "2022-01-23T13:35:33.503000", "content": "", "votes": 1, "replies": [ { "id": 1661473, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-23T13:38:41.500000", "content": "

Idea is really good:

\n", "votes": 1, "replies": [] }, { "id": 1661856, "author_name": "SeshuRaju 🧘‍♂️", "author_url": "", "post_date": "2022-01-23T18:52:36.820000", "content": "

@remekkinas does the same slicing works with 1x images instead of 3x images in training?

", "votes": 0, "replies": [] }, { "id": 1661863, "author_name": "Anirudh", "author_url": "", "post_date": "2022-01-23T19:04:16.857000", "content": "

Training with different resolutions will affect the f2 score badly during prediction.

", "votes": 0, "replies": [] }, { "id": 1661932, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-01-23T20:33:00.527000", "content": "

@anirudhyadav9784 you are right. I think we should train model differently for SAHI. I will test model creation this week. Idea behind this tool is great but it requires time (some experimentation) to achieve good score.

", "votes": 0, "replies": [] }, { "id": 1662283, "author_name": "Anirudh", "author_url": "", "post_date": "2022-01-24T06:55:24.747000", "content": "

Yep tried that but this is what I learned, Don't go beyond 4 slices even then false positive is the real pain. Need to train it much longer with optimized hyp for it to work else its cons will outrank its pros and seems like 30 hours Kaggle GPU limit is not enough for me.

", "votes": 1, "replies": [] } ] }, { "id": 1664087, "author_name": "Parv Yadav", "author_url": "", "post_date": "2022-01-25T17:21:35.260000", "content": "

Interesting work @remekkinas ✨Thanks!😄

", "votes": 1, "replies": [] }, { "id": 1662930, "author_name": "Ismail Hossain Polas", "author_url": "", "post_date": "2022-01-24T16:49:34.870000", "content": "

Thank You for sharing! Interesting!

", "votes": 1, "replies": [] }, { "id": 1662881, "author_name": "Eugene J. Ryu", "author_url": "", "post_date": "2022-01-24T15:55:45.443000", "content": "

Thank you for sharing!!! :D

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1661094": "Today I provided new notebook for experimentations - **SAHI - Slicing Aided Hyper Inference. SAHI is a vision library for large-scale object detection & instance segmentation**. Now I introduced it on Kaggle for Yolov5 (YoloX will be soon - testing it). I think it could be helful for us - we deal with many small object here ....\n\nHope you enjoy it and find it helpful not only in this comeptition but in your work/other competitions.\n\nNotebook: https://www.kaggle.com/remekkinas/sahi-slicing-aided-hyper-inference-yv5-and-yx\n\n\n![video](https://youtu.be/gUex8YoguHU)\n\n\n\n![SAHI](https://user-images.githubusercontent.com/34196005/144092739-c1d9bade-a128-4346-947f-424ce00e5c4f.gif)", "1663537": "Did sahi really work?👀", "1662218": "Thank You For This. Actually i have started learning ML Algorithms from past 15-20 days. But my concepts about what happens in the backend are not getting clear. So if you can guide me on this then it will be very helpful. Or the way you have done ML Algo's Deeply. Basically study material. Once again Thanks alot.", "1662183": "Great ! I am not satisfied with SAHI because it never improved my score\n", "1661777": "@remekkinas Beat me to a Sahi notebook, I am waiting for a good score on my model ;). Great job!", "1661501": "Thanks for sharing. Waiting for yolox!", "1661423": "Thank you for this notebook! And thank you for all your nice sharings during the competition!", "1661195": "Thanks for all the amazing notebooks you have been creating for the competition. I've learned a lot from you. @remekkinas ❤️", "1661191": "I was just trying this on collab... thnx anyways", "1661155": "very nice!", "1662597": "That's really interesting. Amazing work overall @remekkinas ", "1694946": "As far as I can see now SAHI paper is available here: https://arxiv.org/abs/2202.06934", "1694668": "@remekkinas what are your observations after the competition,\nI implemented sliced inference (SAHI) during the competition but it was performing poorly on public leaderboard and thus didn't experiment much on it,however in observations after the competition it gave some large marginal jump in score from public to private leaderboard ( as high as + 0.16 )\nI guess this algorithm has great potential in object detection just needs more research :)", "1661466": "", "1664087": "Interesting work @remekkinas ✨Thanks!😄", "1662930": "Thank You for sharing! Interesting!", "1662881": "Thank you for sharing!!! :D" } }