{ "id": 54307, "title": "Did Allen Draw This in Art Class?", "url": "/competitions/data-science-bowl-2018/discussion/54307", "author_name": "", "post_date": "2018-04-12T03:06:44.046477Z", "votes": 21, "comment_count": 24, "views": 0, "content": "

\"enter

\n\n

\"enter

", "messages": [ { "id": "312605", "postDate": "04/12/2018 03:06:44", "content": "

\"enter

\n\n

\"enter

", "rawMarkdown": "![enter image description here][1]\n\n\n![enter image description here][2]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/312605/9094/5390.png\n\n [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/312605/9106/170b.png", "votes": null }, { "id": "312612", "postDate": "04/12/2018 03:24:15", "content": "

Medical students sketch in the pathology class.

", "rawMarkdown": "Medical students sketch in the pathology class.", "votes": null }, { "id": "312614", "postDate": "04/12/2018 03:37:50", "content": "

BTW it is this one. https://www.kaggle.com/c/data-science-bowl-2018/discussion/47770#312613 It is a TIFF file and not a PNG file.

", "rawMarkdown": "BTW it is this one. https://www.kaggle.com/c/data-science-bowl-2018/discussion/47770#312613 It is a TIFF file and not a PNG file.", "votes": null }, { "id": "312686", "postDate": "04/12/2018 06:36:16", "content": "

Wow!

", "rawMarkdown": "Wow!", "votes": null }, { "id": "312860", "postDate": "04/12/2018 13:27:26", "content": "

Some kind of OCR post-processing would have helped!\n\"enter

", "rawMarkdown": "Some kind of OCR post-processing would have helped!\n![enter image description here][1]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/312860/9124/Fig-5.png", "votes": null }, { "id": "312873", "postDate": "04/12/2018 13:44:10", "content": "

@See-- your results is impressive. is the the results of single model or ensemble?

", "rawMarkdown": "See-- your results is impressive. is the the results of single model or ensemble?", "votes": null }, { "id": "312876", "postDate": "04/12/2018 13:46:21", "content": "

I wish! 🤣

", "rawMarkdown": "I wish! 🤣", "votes": null }, { "id": "312877", "postDate": "04/12/2018 13:47:15", "content": "

I imagine the Fig. 5 won’t matter for most models, but I love that it was included.

", "rawMarkdown": "I imagine the Fig. 5 won’t matter for most models, but I love that it was included.", "votes": null }, { "id": "312879", "postDate": "04/12/2018 13:47:39", "content": "

Thanks! It is a single model. The ensemble submission will take some more time.

", "rawMarkdown": "Thanks! It is a single model. The ensemble submission will take some more time.", "votes": null }, { "id": "312887", "postDate": "04/12/2018 13:58:42", "content": "

@See-- my single model don't detect much. Here is the ensemble results. Your model seems to handle fine detail very well. Seems that mask rcnn cannot handle fine segmentation details (more rounded)

\n\n

\"enter

", "rawMarkdown": "See-- my single model don't detect much. Here is the ensemble results. Your model seems to handle fine detail very well. Seems that mask rcnn cannot handle fine segmentation details (more rounded)\n\n ![enter image description here][1]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/312887/9126/5390acefd575cf9b33413ddf6cbb9ce137ae07dc04616ba24c7b5fe476c827d2.png", "votes": null }, { "id": "312888", "postDate": "04/12/2018 13:59:05", "content": "

This is a very impressive result! I’m speechless.

", "rawMarkdown": "This is a very impressive result! I’m speechless.", "votes": null }, { "id": "312896", "postDate": "04/12/2018 14:07:30", "content": "

Dittos, awesome result See! Maybe nuclei have been embedded in our alphabet all along.

", "rawMarkdown": "Dittos, awesome result See! Maybe nuclei have been embedded in our alphabet all along.", "votes": null }, { "id": "312898", "postDate": "04/12/2018 14:10:35", "content": "

@Heng still very impressive and thanks so much for being so open with insightful comments and great code.

", "rawMarkdown": "Heng still very impressive and thanks so much for being so open with insightful comments and great code.", "votes": null }, { "id": "312902", "postDate": "04/12/2018 14:15:53", "content": "

Wow. Thanks for sharing, @Heng. Did you try using a mask refinement step?

", "rawMarkdown": "Wow. Thanks for sharing, @Heng. Did you try using a mask refinement step?", "votes": null }, { "id": "312907", "postDate": "04/12/2018 14:19:41", "content": "

No. Just clustering and removing false detection. Results is average of clustered predictions. From left to right

\n\n\n\n

final results = cluster - post filtering (too small mask and those at the border of images are removed)

", "rawMarkdown": "No. Just clustering and removing false detection. Results is average of clustered predictions. From left to right\n\n- input image\n\n- final results (color overlay)\n\n- final results (contour overlay)\n\n- cluster(contour overlay)\n\n- cluster(color overlay)\n\n- rejected instances (from ensemble of 12 mask rcnn prediction , flip scale)\n\n- sum of all instances (accepted and rejected)\n\nfinal results = cluster - post filtering (too small mask and those at the border of images are removed)", "votes": null }, { "id": "312916", "postDate": "04/12/2018 14:29:30", "content": "

@ Russ W Thanks! Your public LB score is very impressive. Hope to learn more about it after the competition ends.

", "rawMarkdown": "Russ W Thanks! Your public LB score is very impressive. Hope to learn more about it after the competition ends.", "votes": null }, { "id": "312931", "postDate": "04/12/2018 14:45:11", "content": "

@Allen, gotta confess, strongly motivated by your stratospheric leaderboard score I spent a huge chunk of my time trying to get mask-rcnn working, piecing together code from your repo + great tips and also trying @Heng's pytorch and @Waleed's matterport. Found results a lot more sensitive and less stable than very innovative unet modifications from my teammate Dmytro across some nice synthetic images from teammate bobutis and wonderful extra annotation from teammate Ira. Hope to provide more details later, but in any case, definitely looking forward to learning exactly how you got that LB result!

", "rawMarkdown": "Allen, gotta confess, strongly motivated by your stratospheric leaderboard score I spent a huge chunk of my time trying to get mask-rcnn working, piecing together code from your repo + great tips and also trying @Heng's pytorch and @Waleed's matterport. Found results a lot more sensitive and less stable than very innovative unet modifications from my teammate Dmytro across some nice synthetic images from teammate bobutis and wonderful extra annotation from teammate Ira. Hope to provide more details later, but in any case, definitely looking forward to learning exactly how you got that LB result!", "votes": null }, { "id": "312935", "postDate": "04/12/2018 14:48:34", "content": "

Yeah, I think it’s a trade-off between localization or segmentation error. In my experiments, UNet produced more of the former and less of the later. I also wouldn’t hold my top public score as an indicator of anything useful. It did poorly on phase two.

", "rawMarkdown": "Yeah, I think it’s a trade-off between localization or segmentation error. In my experiments, UNet produced more of the former and less of the later. I also wouldn’t hold my top public score as an indicator of anything useful. It did poorly on phase two.", "votes": null }, { "id": "312946", "postDate": "04/12/2018 15:06:00", "content": "

Nice distinction of the kinds of error, and apppreciate your candidness and modesty :). Stage 2 will carry it's own set of biases, but I still hold your public score in high regard and think your deep experience in the field may have led to some subtle and important design decisions across a huge space of choices. These can manifest and multiply across the whole pipeline: (a) pre-processing, including mask augmentation (b) config parms in mask-rcnn architecture (c) smart post-processing, e.g. nms and which classic morphology techniques work best with modern machine learning outputs vis-a-vis noisy ground truth.

", "rawMarkdown": "Nice distinction of the kinds of error, and apppreciate your candidness and modesty :). Stage 2 will carry it's own set of biases, but I still hold your public score in high regard and think your deep experience in the field may have led to some subtle and important design decisions across a huge space of choices. These can manifest and multiply across the whole pipeline: (a) pre-processing, including mask augmentation (b) config parms in mask-rcnn architecture (c) smart post-processing, e.g. nms and which classic morphology techniques work best with modern machine learning outputs vis-a-vis noisy ground truth.", "votes": null }, { "id": "313257", "postDate": "04/13/2018 02:18:26", "content": "

@Allen，Two question，why score all 0.000？All images in stage2 have object or some are just noise？

", "rawMarkdown": "Allen，Two question，why score all 0.000？All images in stage2 have object or some are just noise？", "votes": null }, { "id": "313705", "postDate": "04/13/2018 18:11:00", "content": "

What the nets I submitted predict on these. These are basically just predictions for nuclei centers, though. Didn't have time to make all of the post-processing stuff. I'd say both your results look great. The major difference is probably just sensitivity. Anyway, the data that had to go into this is probably only going to hurt perf overall.

\n\n

\"enter

\n\n

\"enter

", "rawMarkdown": "What the nets I submitted predict on these. These are basically just predictions for nuclei centers, though. Didn't have time to make all of the post-processing stuff. I'd say both your results look great. The major difference is probably just sensitivity. Anyway, the data that had to go into this is probably only going to hurt perf overall.\n\n![enter image description here][1]\n\n\n![enter image description here][2]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/313705/9170/download%20(1).png\n [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/313705/9171/download%20(7).png", "votes": null }, { "id": "313718", "postDate": "04/13/2018 18:40:43", "content": "

somehow, it suddenly reminds me of the face detection i did about 16 years ago. During that time the breakthrough was adaboost and haar features. Today deep learning has become the popular method.

\n\n

The cmu test data is benchmark data and has a drawn face in its test image. We are surprised that face detection also detects it. http://www.cs.cmu.edu/~har/faces.html

\n\n

Note that work prior to adaboost was neural network face detection.
\nIt is like NN->SVM->adaboost-> deep networks. This summaries the history of face detection in computer vision history

\n\n

\"enter

", "rawMarkdown": "somehow, it suddenly reminds me of the face detection i did about 16 years ago. During that time the breakthrough was adaboost and haar features. Today deep learning has become the popular method.\n\nThe cmu test data is benchmark data and has a drawn face in its test image. We are surprised that face detection also detects it. http://www.cs.cmu.edu/~har/faces.html\n\nNote that work prior to adaboost was neural network face detection. \nIt is like NN->SVM->adaboost-> deep networks. This summaries the history of face detection in computer vision history\n\n ![enter image description here][1]\n\n\n [1]: https://csdl-images.computer.org/trans/tp/2002/01/figures/i003413.gif", "votes": null }, { "id": "314880", "postDate": "04/16/2018 12:27:07", "content": "

I've this one from single model without ensembling, just with normalisation ;) \"enter

", "rawMarkdown": "I've this one from single model without ensembling, just with normalisation ;) ![enter image description here][1]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/314880/9197/figure_test.png", "votes": null }, { "id": "314900", "postDate": "04/16/2018 13:17:58", "content": "

I'm guessing you haven't trained on different zoom levels.

", "rawMarkdown": "I'm guessing you haven't trained on different zoom levels.", "votes": null }, { "id": "314914", "postDate": "04/16/2018 13:49:41", "content": "

Thanks for sharing that process and the output. With a fair bit of openCV experience but limited ML experience, it's super helpful to have that walkthrough.

", "rawMarkdown": "Thanks for sharing that process and the output. With a fair bit of openCV experience but limited ML experience, it's super helpful to have that walkthrough.", "votes": null } ], "comments": [ { "id": 312612, "author_name": "osciiart", "author_url": "", "post_date": "04/12/2018 03:24:15", "content": "

Medical students sketch in the pathology class.

", "votes": null, "replies": [ { "id": 312614, "author_name": "pliptor", "author_url": "", "post_date": "04/12/2018 03:37:50", "content": "

BTW it is this one. https://www.kaggle.com/c/data-science-bowl-2018/discussion/47770#312613 It is a TIFF file and not a PNG file.

", "votes": null, "replies": [] } ] }, { "id": 312686, "author_name": "outliar", "author_url": "", "post_date": "04/12/2018 06:36:16", "content": "

Wow!

", "votes": null, "replies": [] }, { "id": 312860, "author_name": "seesee", "author_url": "", "post_date": "04/12/2018 13:27:26", "content": "

Some kind of OCR post-processing would have helped!\n\"enter

", "votes": null, "replies": [ { "id": 312873, "author_name": "hengck23", "author_url": "", "post_date": "04/12/2018 13:44:10", "content": "

@See-- your results is impressive. is the the results of single model or ensemble?

", "votes": null, "replies": [] }, { "id": 312879, "author_name": "seesee", "author_url": "", "post_date": "04/12/2018 13:47:39", "content": "

Thanks! It is a single model. The ensemble submission will take some more time.

", "votes": null, "replies": [] }, { "id": 312888, "author_name": "agoodman", "author_url": "", "post_date": "04/12/2018 13:59:05", "content": "

This is a very impressive result! I’m speechless.

", "votes": null, "replies": [] }, { "id": 312896, "author_name": "sasrdw", "author_url": "", "post_date": "04/12/2018 14:07:30", "content": "

Dittos, awesome result See! Maybe nuclei have been embedded in our alphabet all along.

", "votes": null, "replies": [] }, { "id": 313257, "author_name": "xhqlcwq", "author_url": "", "post_date": "04/13/2018 02:18:26", "content": "

@Allen，Two question，why score all 0.000？All images in stage2 have object or some are just noise？

", "votes": null, "replies": [] } ] }, { "id": 312876, "author_name": "agoodman", "author_url": "", "post_date": "04/12/2018 13:46:21", "content": "

I wish! 🤣

", "votes": null, "replies": [] }, { "id": 312877, "author_name": "agoodman", "author_url": "", "post_date": "04/12/2018 13:47:15", "content": "

I imagine the Fig. 5 won’t matter for most models, but I love that it was included.

", "votes": null, "replies": [] }, { "id": 312887, "author_name": "hengck23", "author_url": "", "post_date": "04/12/2018 13:58:42", "content": "

@See-- my single model don't detect much. Here is the ensemble results. Your model seems to handle fine detail very well. Seems that mask rcnn cannot handle fine segmentation details (more rounded)

\n\n

\"enter

", "votes": null, "replies": [ { "id": 312898, "author_name": "sasrdw", "author_url": "", "post_date": "04/12/2018 14:10:35", "content": "

@Heng still very impressive and thanks so much for being so open with insightful comments and great code.

", "votes": null, "replies": [] }, { "id": 312902, "author_name": "agoodman", "author_url": "", "post_date": "04/12/2018 14:15:53", "content": "

Wow. Thanks for sharing, @Heng. Did you try using a mask refinement step?

", "votes": null, "replies": [] }, { "id": 312907, "author_name": "hengck23", "author_url": "", "post_date": "04/12/2018 14:19:41", "content": "

No. Just clustering and removing false detection. Results is average of clustered predictions. From left to right

\n\n\n\n

final results = cluster - post filtering (too small mask and those at the border of images are removed)

", "votes": null, "replies": [] }, { "id": 312916, "author_name": "hengck23", "author_url": "", "post_date": "04/12/2018 14:29:30", "content": "

@ Russ W Thanks! Your public LB score is very impressive. Hope to learn more about it after the competition ends.

", "votes": null, "replies": [] }, { "id": 312931, "author_name": "sasrdw", "author_url": "", "post_date": "04/12/2018 14:45:11", "content": "

@Allen, gotta confess, strongly motivated by your stratospheric leaderboard score I spent a huge chunk of my time trying to get mask-rcnn working, piecing together code from your repo + great tips and also trying @Heng's pytorch and @Waleed's matterport. Found results a lot more sensitive and less stable than very innovative unet modifications from my teammate Dmytro across some nice synthetic images from teammate bobutis and wonderful extra annotation from teammate Ira. Hope to provide more details later, but in any case, definitely looking forward to learning exactly how you got that LB result!

", "votes": null, "replies": [] }, { "id": 312935, "author_name": "agoodman", "author_url": "", "post_date": "04/12/2018 14:48:34", "content": "

Yeah, I think it’s a trade-off between localization or segmentation error. In my experiments, UNet produced more of the former and less of the later. I also wouldn’t hold my top public score as an indicator of anything useful. It did poorly on phase two.

", "votes": null, "replies": [] }, { "id": 312946, "author_name": "sasrdw", "author_url": "", "post_date": "04/12/2018 15:06:00", "content": "

Nice distinction of the kinds of error, and apppreciate your candidness and modesty :). Stage 2 will carry it's own set of biases, but I still hold your public score in high regard and think your deep experience in the field may have led to some subtle and important design decisions across a huge space of choices. These can manifest and multiply across the whole pipeline: (a) pre-processing, including mask augmentation (b) config parms in mask-rcnn architecture (c) smart post-processing, e.g. nms and which classic morphology techniques work best with modern machine learning outputs vis-a-vis noisy ground truth.

", "votes": null, "replies": [] }, { "id": 314914, "author_name": "readywater", "author_url": "", "post_date": "04/16/2018 13:49:41", "content": "

Thanks for sharing that process and the output. With a fair bit of openCV experience but limited ML experience, it's super helpful to have that walkthrough.

", "votes": null, "replies": [] } ] }, { "id": 313705, "author_name": "thenickname", "author_url": "", "post_date": "04/13/2018 18:11:00", "content": "

What the nets I submitted predict on these. These are basically just predictions for nuclei centers, though. Didn't have time to make all of the post-processing stuff. I'd say both your results look great. The major difference is probably just sensitivity. Anyway, the data that had to go into this is probably only going to hurt perf overall.

\n\n

\"enter

\n\n

\"enter

", "votes": null, "replies": [] }, { "id": 313718, "author_name": "hengck23", "author_url": "", "post_date": "04/13/2018 18:40:43", "content": "

somehow, it suddenly reminds me of the face detection i did about 16 years ago. During that time the breakthrough was adaboost and haar features. Today deep learning has become the popular method.

\n\n

The cmu test data is benchmark data and has a drawn face in its test image. We are surprised that face detection also detects it. http://www.cs.cmu.edu/~har/faces.html

\n\n

Note that work prior to adaboost was neural network face detection.
\nIt is like NN->SVM->adaboost-> deep networks. This summaries the history of face detection in computer vision history

\n\n

\"enter

", "votes": null, "replies": [] }, { "id": 314880, "author_name": "dawrym", "author_url": "", "post_date": "04/16/2018 12:27:07", "content": "

I've this one from single model without ensembling, just with normalisation ;) \"enter

", "votes": null, "replies": [ { "id": 314900, "author_name": "thenickname", "author_url": "", "post_date": "04/16/2018 13:17:58", "content": "

I'm guessing you haven't trained on different zoom levels.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "312605": "![enter image description here][1]\n\n\n![enter image description here][2]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/312605/9094/5390.png\n\n [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/312605/9106/170b.png", "312612": "Medical students sketch in the pathology class.", "312614": "BTW it is this one. https://www.kaggle.com/c/data-science-bowl-2018/discussion/47770#312613 It is a TIFF file and not a PNG file.", "312686": "Wow!", "312860": "Some kind of OCR post-processing would have helped!\n![enter image description here][1]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/312860/9124/Fig-5.png", "312873": "See-- your results is impressive. is the the results of single model or ensemble?", "312876": "I wish! 🤣", "312877": "I imagine the Fig. 5 won’t matter for most models, but I love that it was included.", "312879": "Thanks! It is a single model. The ensemble submission will take some more time.", "312887": "See-- my single model don't detect much. Here is the ensemble results. Your model seems to handle fine detail very well. Seems that mask rcnn cannot handle fine segmentation details (more rounded)\n\n ![enter image description here][1]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/312887/9126/5390acefd575cf9b33413ddf6cbb9ce137ae07dc04616ba24c7b5fe476c827d2.png", "312888": "This is a very impressive result! I’m speechless.", "312896": "Dittos, awesome result See! Maybe nuclei have been embedded in our alphabet all along.", "312898": "Heng still very impressive and thanks so much for being so open with insightful comments and great code.", "312902": "Wow. Thanks for sharing, @Heng. Did you try using a mask refinement step?", "312907": "No. Just clustering and removing false detection. Results is average of clustered predictions. From left to right\n\n- input image\n\n- final results (color overlay)\n\n- final results (contour overlay)\n\n- cluster(contour overlay)\n\n- cluster(color overlay)\n\n- rejected instances (from ensemble of 12 mask rcnn prediction , flip scale)\n\n- sum of all instances (accepted and rejected)\n\nfinal results = cluster - post filtering (too small mask and those at the border of images are removed)", "312916": "Russ W Thanks! Your public LB score is very impressive. Hope to learn more about it after the competition ends.", "312931": "Allen, gotta confess, strongly motivated by your stratospheric leaderboard score I spent a huge chunk of my time trying to get mask-rcnn working, piecing together code from your repo + great tips and also trying @Heng's pytorch and @Waleed's matterport. Found results a lot more sensitive and less stable than very innovative unet modifications from my teammate Dmytro across some nice synthetic images from teammate bobutis and wonderful extra annotation from teammate Ira. Hope to provide more details later, but in any case, definitely looking forward to learning exactly how you got that LB result!", "312935": "Yeah, I think it’s a trade-off between localization or segmentation error. In my experiments, UNet produced more of the former and less of the later. I also wouldn’t hold my top public score as an indicator of anything useful. It did poorly on phase two.", "312946": "Nice distinction of the kinds of error, and apppreciate your candidness and modesty :). Stage 2 will carry it's own set of biases, but I still hold your public score in high regard and think your deep experience in the field may have led to some subtle and important design decisions across a huge space of choices. These can manifest and multiply across the whole pipeline: (a) pre-processing, including mask augmentation (b) config parms in mask-rcnn architecture (c) smart post-processing, e.g. nms and which classic morphology techniques work best with modern machine learning outputs vis-a-vis noisy ground truth.", "313257": "Allen，Two question，why score all 0.000？All images in stage2 have object or some are just noise？", "313705": "What the nets I submitted predict on these. These are basically just predictions for nuclei centers, though. Didn't have time to make all of the post-processing stuff. I'd say both your results look great. The major difference is probably just sensitivity. Anyway, the data that had to go into this is probably only going to hurt perf overall.\n\n![enter image description here][1]\n\n\n![enter image description here][2]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/313705/9170/download%20(1).png\n [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/313705/9171/download%20(7).png", "313718": "somehow, it suddenly reminds me of the face detection i did about 16 years ago. During that time the breakthrough was adaboost and haar features. Today deep learning has become the popular method.\n\nThe cmu test data is benchmark data and has a drawn face in its test image. We are surprised that face detection also detects it. http://www.cs.cmu.edu/~har/faces.html\n\nNote that work prior to adaboost was neural network face detection. \nIt is like NN->SVM->adaboost-> deep networks. This summaries the history of face detection in computer vision history\n\n ![enter image description here][1]\n\n\n [1]: https://csdl-images.computer.org/trans/tp/2002/01/figures/i003413.gif", "314880": "I've this one from single model without ensembling, just with normalisation ;) ![enter image description here][1]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/314880/9197/figure_test.png", "314900": "I'm guessing you haven't trained on different zoom levels.", "314914": "Thanks for sharing that process and the output. With a fair bit of openCV experience but limited ML experience, it's super helpful to have that walkthrough." }, "source": "meta" }