![\"enter](\"https://news.developer.nvidia.com/wp-content/uploads/2016/01/Diagnosing-Cancer-with-Deep-Learning-and-GPUs.png\")
Anyone would be using instance segmentation like mask rcnn? if you do, share your results and discussion here!
", "messages": [ { "id": "270087", "postDate": "01/17/2018 18:38:33", "content": "Anyone would be using instance segmentation like mask rcnn? if you do, share your results and discussion here!
", "rawMarkdown": "Anyone would be using instance segmentation like mask rcnn? if you do, share your results and discussion here!", "votes": null }, { "id": "270120", "postDate": "01/17/2018 19:15:16", "content": "Hi, I’m an organizer, but I’m planning to participate for fun (i.e. I’m not eligible for a prize). I also work on Keras-RCNN which includes Mask-RCNN functionality and plan on using it for the competition. I created a GitHub pull request to track DSB-related stuff. It’d be awesome if other people contributed!
", "rawMarkdown": "Hi, I’m an organizer, but I’m planning to participate for fun (i.e. I’m not eligible for a prize). I also work on [Keras-RCNN][1] which includes Mask-RCNN functionality and plan on using it for the competition. I created a [GitHub pull request][2] to track DSB-related stuff. It’d be awesome if other people contributed!\n\n [1]: https://github.com/broadinstitute/keras-rcnn\n [2]: https://github.com/broadinstitute/keras-rcnn/pull/152", "votes": null }, { "id": "270939", "postDate": "01/19/2018 09:42:57", "content": "https://news.developer.nvidia.com/diagnosing-cancer-with-deep-learning-and-gpus/
\n\n2015 Gland Segmentation Challenge solution
\n\n
\"Deep Watershed Transformation for Instance Segmentation\" - Min Bai and Raquel Urtasun, in CVPR 2017.\nhttps://arxiv.org/abs/1611.08303.
\n\n\"End-to-End Instance Segmentation with Recurrent Attention\" - Mengye Ren, Richard S. Zemel, CVPR 2017
\n\nmany more:
\n\n\n\nhttp://www.bioimageanalysis.org/wp/wp-content/uploads/formidable/6/MIMO-Net_Isbi2017.pdf
\n\nhttps://arxiv.org/pdf/1607.04889.pdf
\n\n", "rawMarkdown": "many more:\n\nhttp://openaccess.thecvf.com/content_ICCV_2017/papers/Liu_SGN_Sequential_Grouping_ICCV_2017_paper.pdf\n\nhttp://www.bioimageanalysis.org/wp/wp-content/uploads/formidable/6/MIMO-Net_Isbi2017.pdf\n\nhttps://arxiv.org/pdf/1607.04889.pdf\n\nhttps://www.researchgate.net/profile/Yizhe_Zhang13/publication/319460791_Deep_Adversarial_Networks_for_Biomedical_Image_Segmentation_Utilizing_Unannotated_Images/links/59c9163fa6fdccc71929d61b/Deep-Adversarial-Networks-for-Biomedical-Image-Segmentation-Utilizing-Unannotated-Images.pdf", "votes": null }, { "id": "270974", "postDate": "01/19/2018 11:08:32", "content": "a proposed solution: learn 3-class: seed, foreground and background\n ![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/270974/8308/instance_seg.png\")
another proposed solution is to treat contour as background
\n\n![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/270974/8314/contour.png\")
I really like that gland segmentation approach, but how to implement it?
", "rawMarkdown": "I really like that gland segmentation approach, but how to implement it?", "votes": null }, { "id": "271075", "postDate": "01/19/2018 16:26:02", "content": "Found some implementation: https://github.com/lisjin/dcan-tensorflow
", "rawMarkdown": "Found some implementation: https://github.com/lisjin/dcan-tensorflow", "votes": null }, { "id": "271272", "postDate": "01/20/2018 02:01:38", "content": "check this too: https://data.broadinstitute.org/bbbc/BBBC006/
", "rawMarkdown": "check this too: https://data.broadinstitute.org/bbbc/BBBC006/", "votes": null }, { "id": "271340", "postDate": "01/20/2018 07:42:47", "content": "https://surabhibhargava.github.io/Detection-and-Segmentation-of-Nuclei-in-Computational-Pathology/
\n\n![\"enter](\"https://surabhibhargava.github.io/assets/arch.jpg\")
![\"enter](\"https://surabhibhargava.github.io/assets/postpro.jpg\")
![\"enter](\"https://surabhibhargava.github.io/assets/3c.png\")
https://www.youtube.com/watch?v=4FhdkiZ51Js
\n\nP. Kainz, M. Urschler, S. Schulter, P. Wohlhart, and V. Lepetit. You should use regression to detect cells. In A. F. Frangi, J. Hornegger, N. Navab, and S. Wells, editors, Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Lecture Notes in Computer Science, Munich, Germany, Oct. 5-9 2015. Springer.
\n\n", "rawMarkdown": "https://www.youtube.com/watch?v=4FhdkiZ51Js\n\nP. Kainz, M. Urschler, S. Schulter, P. Wohlhart, and V. Lepetit. You should use regression to detect cells. In A. F. Frangi, J. Hornegger, N. Navab, and S. Wells, editors, Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Lecture Notes in Computer Science, Munich, Germany, Oct. 5-9 2015. Springer.\n\nhttps://www.tugraz.at/fileadmin/user_upload/Institute/ICG/Images/team_lepetit/publications/kainz_miccai15.pdf", "votes": null }, { "id": "272017", "postDate": "01/22/2018 04:11:53", "content": "https://arxiv.org/pdf/1709.05932.pdf\nhttps://arxiv.org/ftp/arxiv/papers/1709/1709.00179.pdf
\n\n![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/272017/8334/dt.png\")
Facebook just open sourced the code and model for the mask-rcnn: https://github.com/facebookresearch/Detectron
", "rawMarkdown": "Facebook just open sourced the code and model for the mask-rcnn: https://github.com/facebookresearch/Detectron", "votes": null }, { "id": "272767", "postDate": "01/23/2018 16:57:53", "content": "FWIW, this is – more or less – my current approach. I'm predicting foreground and edges, and then using the edges to guide segmentation. It performs pretty well: 0.411 on the LB from a single model. The fiddly details of the segmentation seem to be at least as important as the base neural network to making this work though.
", "rawMarkdown": "FWIW, this is – more or less – my current approach. I'm predicting foreground and edges, and then using the edges to guide segmentation. It performs pretty well: 0.411 on the LB from a single model. The fiddly details of the segmentation seem to be at least as important as the base neural network to making this work though.", "votes": null }, { "id": "272906", "postDate": "01/24/2018 00:03:12", "content": "Hi, if I understand correctly, once we have the edges, we already have the segmentation. So may I ask what you mean by the last sentence, do we need to refine the segments from the edges?
", "rawMarkdown": "Hi, if I understand correctly, once we have the edges, we already have the segmentation. So may I ask what you mean by the last sentence, do we need to refine the segments from the edges?", "votes": null }, { "id": "273004", "postDate": "01/24/2018 02:27:41", "content": "What does FWIW means?
", "rawMarkdown": "What does FWIW means?", "votes": null }, { "id": "273347", "postDate": "01/24/2018 13:41:31", "content": "Nevermind I was able to figure out what FWIW is. Thanks.
", "rawMarkdown": "Nevermind I was able to figure out what FWIW is. Thanks.", "votes": null }, { "id": "273640", "postDate": "01/24/2018 22:29:07", "content": "@Yan Wang, In my case at least, I'm predicting the edges as a bit mask. So I end up with one, approximate, bit mask of the edges. And another of the foreground. Just having the bitmap of the edges doesn't automatically give me the masks for each nuclei, I still have to split them up. And that is the fiddly part.
", "rawMarkdown": "Yan Wang, In my case at least, I'm predicting the edges as a bit mask. So I end up with one, approximate, bit mask of the edges. And another of the foreground. Just having the bitmap of the edges doesn't automatically give me the masks for each nuclei, I still have to split them up. And that is the fiddly part.", "votes": null }, { "id": "273641", "postDate": "01/24/2018 22:29:59", "content": "@WilliamGreen, glad you figured it out. It's one of those very technical abbreviations ;-)
", "rawMarkdown": "WilliamGreen, glad you figured it out. It's one of those very technical abbreviations ;-)", "votes": null }, { "id": "274656", "postDate": "01/27/2018 02:04:49", "content": "Hi, Tim Hochberg, I get the boundary through mask minus erosion in 1pixel or then dilate them in 2-3 pixels, but I try many deep models it always hardly to find the boundary..Do u have any idea?
", "rawMarkdown": "Hi, Tim Hochberg, I get the boundary through mask minus erosion in 1pixel or then dilate them in 2-3 pixels, but I try many deep models it always hardly to find the boundary..Do u have any idea?", "votes": null }, { "id": "274850", "postDate": "01/27/2018 14:39:52", "content": "@bonsen, I'm actually using between 4 and 10 pixel wide edges (still playing around), so perhaps just try using larger edges. Also when I'm selecting crops to feed to the net at training time, I arrange to give it crops that contain an edge at least half the time.
", "rawMarkdown": "bonsen, I'm actually using between 4 and 10 pixel wide edges (still playing around), so perhaps just try using larger edges. Also when I'm selecting crops to feed to the net at training time, I arrange to give it crops that contain an edge at least half the time.", "votes": null }, { "id": "276585", "postDate": "01/31/2018 20:39:47", "content": "For detectron you need a GPU with RAM 4GB or more. I just went thru installing it thru docker and with my 2GB GPU, I couldn't even able to run inference demo. Probably for training you may need multiple GPUs.
", "rawMarkdown": "For detectron you need a GPU with RAM 4GB or more. I just went thru installing it thru docker and with my 2GB GPU, I couldn't even able to run inference demo. Probably for training you may need multiple GPUs.", "votes": null }, { "id": "277020", "postDate": "02/02/2018 02:21:17", "content": "check this latest paper!\nhttps://arxiv.org/pdf/1802.00383.pdf
\n\n![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/277020/8471/2.png\")
![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/277020/8470/1.png\")
When u get the seed and the segmentation of forgorund, how to mix them for different colors?
", "rawMarkdown": "When u get the seed and the segmentation of forgorund, how to mix them for different colors?", "votes": null }, { "id": "277088", "postDate": "02/02/2018 05:23:35", "content": "I will try, thanks.
", "rawMarkdown": "I will try, thanks.", "votes": null }, { "id": "277169", "postDate": "02/02/2018 10:56:49", "content": "Hi, Tim Hochberg, \nI'm sorry that I come to bother u again..... \nIf I use wide edges to train and I will get the wide edges after prediction. It seems that I can't use it to guide the segmentation.\nThanks
", "rawMarkdown": "Hi, Tim Hochberg, \nI'm sorry that I come to bother u again..... \nIf I use wide edges to train and I will get the wide edges after prediction. It seems that I can't use it to guide the segmentation.\nThanks", "votes": null }, { "id": "277596", "postDate": "02/03/2018 17:03:09", "content": "I don't think the stated results in paper is easy to reproduce. Feel too good to be true.
", "rawMarkdown": "I don't think the stated results in paper is easy to reproduce. Feel too good to be true.", "votes": null }, { "id": "277761", "postDate": "02/04/2018 05:57:42", "content": "@Heng CherKeng: What is your best result using mask rcnn? Thanks
", "rawMarkdown": "Heng CherKeng: What is your best result using mask rcnn? Thanks", "votes": null }, { "id": "277765", "postDate": "02/04/2018 06:23:43", "content": "i haven't made a mask-rcnn submission yet.
\n\ni am fine-tunning my code, and now only doing training/validation on gray images (and not the color ones)
\n\non local validation set, i have score of about 0.65. (from previous results, this validation set is about 0.1 higher than LB score)
", "rawMarkdown": "i haven't made a mask-rcnn submission yet. \n\ni am fine-tunning my code, and now only doing training/validation on gray images (and not the color ones)\n\non local validation set, i have score of about 0.65. (from previous results, this validation set is about 0.1 higher than LB score)", "votes": null }, { "id": "277804", "postDate": "02/04/2018 11:45:23", "content": "Great. I think your score will be better when you split data into some clusters (3 to 5) as someone did. However, it must create 3/5 networks which used for each cluster
", "rawMarkdown": "Great. I think your score will be better when you split data into some clusters (3 to 5) as someone did. However, it must create 3/5 networks which used for each cluster", "votes": null }, { "id": "277808", "postDate": "02/04/2018 12:01:48", "content": "i just posted some visual results at my thread:
\n\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/47686
\n\nsee \"i try to write a pytorch mask-rcnn from scratch. ....\nHere are some results from rpn(red is before nms, white is after) and rcnn (pink).\"
", "rawMarkdown": "i just posted some visual results at my thread:\n\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/47686\n\nsee \"i try to write a pytorch mask-rcnn from scratch. ....\nHere are some results from rpn(red is before nms, white is after) and rcnn (pink).\"", "votes": null }, { "id": "277853", "postDate": "02/04/2018 16:06:40", "content": "when u extract the object pixels . should i try to seperate them to many images like in the dataset, is that neceesary?
", "rawMarkdown": "when u extract the object pixels . should i try to seperate them to many images like in the dataset, is that neceesary?", "votes": null }, { "id": "278569", "postDate": "02/06/2018 13:54:51", "content": "Hi Ertuka. Do you use Detectron for this dataset. It must be convert to COCO format. Do you have any idea to do it? Thanks
", "rawMarkdown": "Hi Ertuka. Do you use Detectron for this dataset. It must be convert to COCO format. Do you have any idea to do it? Thanks", "votes": null }, { "id": "279496", "postDate": "02/08/2018 05:18:14", "content": "@john1231983 I have failed to be able to run Detectron demo on my GPU since requires at least 4GB RAM so never tried it further. Possibly one may more than 6GB for any chance to apply to this
", "rawMarkdown": "john1231983 I have failed to be able to run Detectron demo on my GPU since requires at least 4GB RAM so never tried it further. Possibly one may more than 6GB for any chance to apply to this", "votes": null }, { "id": "279509", "postDate": "02/08/2018 05:59:21", "content": "Hi @Heng, This article looks very promising. Have you tried it?
", "rawMarkdown": "Hi @Heng, This article looks very promising. Have you tried it?", "votes": null }, { "id": "279690", "postDate": "02/08/2018 14:41:46", "content": "I generated the mask inside objects and contours with unet. But I don't know how to use the contour to split different objects. Can you describe your edge guided segmentation method? Thanks.
", "rawMarkdown": "I generated the mask inside objects and contours with unet. But I don't know how to use the contour to split different objects. Can you describe your edge guided segmentation method? Thanks.", "votes": null }, { "id": "279721", "postDate": "02/08/2018 15:43:29", "content": "@zmingen, I use something pretty similar to what's described here: http://scikit-image.org/docs/dev/auto_examples/segmentation/plot_watershed.html. I use the background - edges to generate seeds, and then to watershed segmentation to fill out from the seeds.
", "rawMarkdown": "zmingen, I use something pretty similar to what's described here: http://scikit-image.org/docs/dev/auto_examples/segmentation/plot_watershed.html. I use the background - edges to generate seeds, and then to watershed segmentation to fill out from the seeds.", "votes": null }, { "id": "284980", "postDate": "02/19/2018 01:24:41", "content": "this paper has better results than mask rcnn:
\n\nMaskLab: Instance Segmentation by Refining Object Detection with Semantic and Direction Features
\n\n", "rawMarkdown": "this paper has better results than mask rcnn:\n\nMaskLab: Instance Segmentation by Refining Object Detection with Semantic and Direction Features\n\nhttp://liangchiehchen.com/", "votes": null }, { "id": "286308", "postDate": "02/22/2018 00:56:20", "content": "but they have not released the code yet!
", "rawMarkdown": "but they have not released the code yet!", "votes": null }, { "id": "1370634", "postDate": "06/30/2021 09:38:29", "content": "Hey! Really faultless Mask RCNN installation for TensorFlow GPU
\n 👉 https://github.com/BuseYaren/Installation-MaskRCNN
Hey! Really faultless Mask RCNN installation for TensorFlow GPU
\n 👉 https://github.com/BuseYaren/Installation-MaskRCNN
Hi, I’m an organizer, but I’m planning to participate for fun (i.e. I’m not eligible for a prize). I also work on Keras-RCNN which includes Mask-RCNN functionality and plan on using it for the competition. I created a GitHub pull request to track DSB-related stuff. It’d be awesome if other people contributed!
", "votes": null, "replies": [] }, { "id": 270939, "author_name": "hengck23", "author_url": "", "post_date": "01/19/2018 09:42:57", "content": "https://news.developer.nvidia.com/diagnosing-cancer-with-deep-learning-and-gpus/
\n\n2015 Gland Segmentation Challenge solution
\n\n
I really like that gland segmentation approach, but how to implement it?
", "votes": null, "replies": [] }, { "id": 271075, "author_name": "serhiy", "author_url": "", "post_date": "01/19/2018 16:26:02", "content": "Found some implementation: https://github.com/lisjin/dcan-tensorflow
", "votes": null, "replies": [] }, { "id": 271272, "author_name": "hengck23", "author_url": "", "post_date": "01/20/2018 02:01:38", "content": "check this too: https://data.broadinstitute.org/bbbc/BBBC006/
", "votes": null, "replies": [] }, { "id": 272767, "author_name": "bitsofbits", "author_url": "", "post_date": "01/23/2018 16:57:53", "content": "FWIW, this is – more or less – my current approach. I'm predicting foreground and edges, and then using the edges to guide segmentation. It performs pretty well: 0.411 on the LB from a single model. The fiddly details of the segmentation seem to be at least as important as the base neural network to making this work though.
", "votes": null, "replies": [] }, { "id": 272906, "author_name": "wayne999", "author_url": "", "post_date": "01/24/2018 00:03:12", "content": "Hi, if I understand correctly, once we have the edges, we already have the segmentation. So may I ask what you mean by the last sentence, do we need to refine the segments from the edges?
", "votes": null, "replies": [] }, { "id": 273004, "author_name": "dskswu", "author_url": "", "post_date": "01/24/2018 02:27:41", "content": "What does FWIW means?
", "votes": null, "replies": [] }, { "id": 273347, "author_name": "dskswu", "author_url": "", "post_date": "01/24/2018 13:41:31", "content": "Nevermind I was able to figure out what FWIW is. Thanks.
", "votes": null, "replies": [] }, { "id": 273640, "author_name": "bitsofbits", "author_url": "", "post_date": "01/24/2018 22:29:07", "content": "@Yan Wang, In my case at least, I'm predicting the edges as a bit mask. So I end up with one, approximate, bit mask of the edges. And another of the foreground. Just having the bitmap of the edges doesn't automatically give me the masks for each nuclei, I still have to split them up. And that is the fiddly part.
", "votes": null, "replies": [] }, { "id": 273641, "author_name": "bitsofbits", "author_url": "", "post_date": "01/24/2018 22:29:59", "content": "@WilliamGreen, glad you figured it out. It's one of those very technical abbreviations ;-)
", "votes": null, "replies": [] }, { "id": 274656, "author_name": "bonsen", "author_url": "", "post_date": "01/27/2018 02:04:49", "content": "Hi, Tim Hochberg, I get the boundary through mask minus erosion in 1pixel or then dilate them in 2-3 pixels, but I try many deep models it always hardly to find the boundary..Do u have any idea?
", "votes": null, "replies": [] }, { "id": 274850, "author_name": "bitsofbits", "author_url": "", "post_date": "01/27/2018 14:39:52", "content": "@bonsen, I'm actually using between 4 and 10 pixel wide edges (still playing around), so perhaps just try using larger edges. Also when I'm selecting crops to feed to the net at training time, I arrange to give it crops that contain an edge at least half the time.
", "votes": null, "replies": [] }, { "id": 277088, "author_name": "bonsen", "author_url": "", "post_date": "02/02/2018 05:23:35", "content": "I will try, thanks.
", "votes": null, "replies": [] }, { "id": 277169, "author_name": "bonsen", "author_url": "", "post_date": "02/02/2018 10:56:49", "content": "Hi, Tim Hochberg, \nI'm sorry that I come to bother u again..... \nIf I use wide edges to train and I will get the wide edges after prediction. It seems that I can't use it to guide the segmentation.\nThanks
", "votes": null, "replies": [] }, { "id": 279690, "author_name": "zmingen", "author_url": "", "post_date": "02/08/2018 14:41:46", "content": "I generated the mask inside objects and contours with unet. But I don't know how to use the contour to split different objects. Can you describe your edge guided segmentation method? Thanks.
", "votes": null, "replies": [] }, { "id": 279721, "author_name": "bitsofbits", "author_url": "", "post_date": "02/08/2018 15:43:29", "content": "@zmingen, I use something pretty similar to what's described here: http://scikit-image.org/docs/dev/auto_examples/segmentation/plot_watershed.html. I use the background - edges to generate seeds, and then to watershed segmentation to fill out from the seeds.
", "votes": null, "replies": [] } ] }, { "id": 270940, "author_name": "hengck23", "author_url": "", "post_date": "01/19/2018 09:45:10", "content": "\"Deep Watershed Transformation for Instance Segmentation\" - Min Bai and Raquel Urtasun, in CVPR 2017.\nhttps://arxiv.org/abs/1611.08303.
\n\nHi @Heng, This article looks very promising. Have you tried it?
", "votes": null, "replies": [] } ] }, { "id": 270944, "author_name": "hengck23", "author_url": "", "post_date": "01/19/2018 09:55:14", "content": "\"End-to-End Instance Segmentation with Recurrent Attention\" - Mengye Ren, Richard S. Zemel, CVPR 2017
\n\nmany more:
\n\n\n\nhttp://www.bioimageanalysis.org/wp/wp-content/uploads/formidable/6/MIMO-Net_Isbi2017.pdf
\n\nhttps://arxiv.org/pdf/1607.04889.pdf
\n\n", "votes": null, "replies": [] }, { "id": 270974, "author_name": "hengck23", "author_url": "", "post_date": "01/19/2018 11:08:32", "content": "a proposed solution: learn 3-class: seed, foreground and background\n
another proposed solution is to treat contour as background
\n\n
https://surabhibhargava.github.io/Detection-and-Segmentation-of-Nuclei-in-Computational-Pathology/
\n\n
When u get the seed and the segmentation of forgorund, how to mix them for different colors?
", "votes": null, "replies": [] } ] }, { "id": 271681, "author_name": "hengck23", "author_url": "", "post_date": "01/21/2018 05:47:14", "content": "https://www.youtube.com/watch?v=4FhdkiZ51Js
\n\nP. Kainz, M. Urschler, S. Schulter, P. Wohlhart, and V. Lepetit. You should use regression to detect cells. In A. F. Frangi, J. Hornegger, N. Navab, and S. Wells, editors, Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Lecture Notes in Computer Science, Munich, Germany, Oct. 5-9 2015. Springer.
\n\n", "votes": null, "replies": [] }, { "id": 272017, "author_name": "hengck23", "author_url": "", "post_date": "01/22/2018 04:11:53", "content": "https://arxiv.org/pdf/1709.05932.pdf\nhttps://arxiv.org/ftp/arxiv/papers/1709/1709.00179.pdf
\n\n
Facebook just open sourced the code and model for the mask-rcnn: https://github.com/facebookresearch/Detectron
", "votes": null, "replies": [ { "id": 276585, "author_name": "esorar", "author_url": "", "post_date": "01/31/2018 20:39:47", "content": "For detectron you need a GPU with RAM 4GB or more. I just went thru installing it thru docker and with my 2GB GPU, I couldn't even able to run inference demo. Probably for training you may need multiple GPUs.
", "votes": null, "replies": [] }, { "id": 278569, "author_name": "john1231983", "author_url": "", "post_date": "02/06/2018 13:54:51", "content": "Hi Ertuka. Do you use Detectron for this dataset. It must be convert to COCO format. Do you have any idea to do it? Thanks
", "votes": null, "replies": [] }, { "id": 279496, "author_name": "esorar", "author_url": "", "post_date": "02/08/2018 05:18:14", "content": "@john1231983 I have failed to be able to run Detectron demo on my GPU since requires at least 4GB RAM so never tried it further. Possibly one may more than 6GB for any chance to apply to this
", "votes": null, "replies": [] } ] }, { "id": 277020, "author_name": "hengck23", "author_url": "", "post_date": "02/02/2018 02:21:17", "content": "check this latest paper!\nhttps://arxiv.org/pdf/1802.00383.pdf
\n\n
I don't think the stated results in paper is easy to reproduce. Feel too good to be true.
", "votes": null, "replies": [] } ] }, { "id": 277761, "author_name": "john1231983", "author_url": "", "post_date": "02/04/2018 05:57:42", "content": "@Heng CherKeng: What is your best result using mask rcnn? Thanks
", "votes": null, "replies": [ { "id": 277765, "author_name": "hengck23", "author_url": "", "post_date": "02/04/2018 06:23:43", "content": "i haven't made a mask-rcnn submission yet.
\n\ni am fine-tunning my code, and now only doing training/validation on gray images (and not the color ones)
\n\non local validation set, i have score of about 0.65. (from previous results, this validation set is about 0.1 higher than LB score)
", "votes": null, "replies": [] }, { "id": 277804, "author_name": "john1231983", "author_url": "", "post_date": "02/04/2018 11:45:23", "content": "Great. I think your score will be better when you split data into some clusters (3 to 5) as someone did. However, it must create 3/5 networks which used for each cluster
", "votes": null, "replies": [] }, { "id": 277808, "author_name": "hengck23", "author_url": "", "post_date": "02/04/2018 12:01:48", "content": "i just posted some visual results at my thread:
\n\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/47686
\n\nsee \"i try to write a pytorch mask-rcnn from scratch. ....\nHere are some results from rpn(red is before nms, white is after) and rcnn (pink).\"
", "votes": null, "replies": [] } ] }, { "id": 277853, "author_name": "aviwolfson", "author_url": "", "post_date": "02/04/2018 16:06:40", "content": "when u extract the object pixels . should i try to seperate them to many images like in the dataset, is that neceesary?
", "votes": null, "replies": [] }, { "id": 284980, "author_name": "hengck23", "author_url": "", "post_date": "02/19/2018 01:24:41", "content": "this paper has better results than mask rcnn:
\n\nMaskLab: Instance Segmentation by Refining Object Detection with Semantic and Direction Features
\n\n", "votes": null, "replies": [ { "id": 286308, "author_name": "maghi001", "author_url": "", "post_date": "02/22/2018 00:56:20", "content": "but they have not released the code yet!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "270087": "Anyone would be using instance segmentation like mask rcnn? if you do, share your results and discussion here!", "270120": "Hi, I’m an organizer, but I’m planning to participate for fun (i.e. I’m not eligible for a prize). I also work on [Keras-RCNN][1] which includes Mask-RCNN functionality and plan on using it for the competition. I created a [GitHub pull request][2] to track DSB-related stuff. It’d be awesome if other people contributed!\n\n [1]: https://github.com/broadinstitute/keras-rcnn\n [2]: https://github.com/broadinstitute/keras-rcnn/pull/152", "270939": "https://news.developer.nvidia.com/diagnosing-cancer-with-deep-learning-and-gpus/\n\n 2015 Gland Segmentation Challenge solution\n\n ![enter image description here][1]\n\n\n [1]: https://news.developer.nvidia.com/wp-content/uploads/2016/01/Diagnosing-Cancer-with-Deep-Learning-and-GPUs.png", "270940": "\"Deep Watershed Transformation for Instance Segmentation\" - Min Bai and Raquel Urtasun, in CVPR 2017.\nhttps://arxiv.org/abs/1611.08303.\n\n ![enter image description here][1]\n\n ![enter image description here][2]\n\n ![enter image description here][3]\n \n\n\n [1]: http://img.blog.csdn.net/20170702200353295?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYmVhX3RyZWU=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\n\n [2]: http://img.blog.csdn.net/20170703222133261?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYmVhX3RyZWU=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\n\n [3]: http://img.blog.csdn.net/20170703223354397?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYmVhX3RyZWU=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast", "270944": "\"End-to-End Instance Segmentation with Recurrent Attention\" - Mengye Ren, Richard S. Zemel, CVPR 2017\n\n ![enter image description here][1]\n\n ![enter image description here][2]\n\n ![enter image description here][3]\n\n\n [1]: http://img.blog.csdn.net/20171016102609952?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\n [2]: http://img.blog.csdn.net/20171016101725562?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\n [3]: http://img.blog.csdn.net/20171016101647474?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast", "270951": "many more:\n\nhttp://openaccess.thecvf.com/content_ICCV_2017/papers/Liu_SGN_Sequential_Grouping_ICCV_2017_paper.pdf\n\nhttp://www.bioimageanalysis.org/wp/wp-content/uploads/formidable/6/MIMO-Net_Isbi2017.pdf\n\nhttps://arxiv.org/pdf/1607.04889.pdf\n\nhttps://www.researchgate.net/profile/Yizhe_Zhang13/publication/319460791_Deep_Adversarial_Networks_for_Biomedical_Image_Segmentation_Utilizing_Unannotated_Images/links/59c9163fa6fdccc71929d61b/Deep-Adversarial-Networks-for-Biomedical-Image-Segmentation-Utilizing-Unannotated-Images.pdf", "270974": "a proposed solution: learn 3-class: seed, foreground and background\n ![enter image description here][1]\n\nanother proposed solution is to treat contour as background\n\n ![enter image description here][2]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/270974/8308/instance_seg.png\n [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/270974/8314/contour.png", "271072": "I really like that gland segmentation approach, but how to implement it?", "271075": "Found some implementation: https://github.com/lisjin/dcan-tensorflow", "271272": "check this too: https://data.broadinstitute.org/bbbc/BBBC006/", "271340": "https://surabhibhargava.github.io/Detection-and-Segmentation-of-Nuclei-in-Computational-Pathology/\n\n ![enter image description here][1]\n\n ![enter image description here][2]\n\n ![enter image description here][3]\n\n\n [1]: https://surabhibhargava.github.io/assets/arch.jpg\n [2]: https://surabhibhargava.github.io/assets/postpro.jpg\n [3]: https://surabhibhargava.github.io/assets/3c.png", "271681": "https://www.youtube.com/watch?v=4FhdkiZ51Js\n\nP. Kainz, M. Urschler, S. Schulter, P. Wohlhart, and V. Lepetit. You should use regression to detect cells. In A. F. Frangi, J. Hornegger, N. Navab, and S. Wells, editors, Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Lecture Notes in Computer Science, Munich, Germany, Oct. 5-9 2015. Springer.\n\nhttps://www.tugraz.at/fileadmin/user_upload/Institute/ICG/Images/team_lepetit/publications/kainz_miccai15.pdf", "272017": "https://arxiv.org/pdf/1709.05932.pdf\nhttps://arxiv.org/ftp/arxiv/papers/1709/1709.00179.pdf\n\n ![enter image description here][1]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/272017/8334/dt.png", "272704": "Facebook just open sourced the code and model for the mask-rcnn: https://github.com/facebookresearch/Detectron", "272767": "FWIW, this is – more or less – my current approach. I'm predicting foreground and edges, and then using the edges to guide segmentation. It performs pretty well: 0.411 on the LB from a single model. The fiddly details of the segmentation seem to be at least as important as the base neural network to making this work though.", "272906": "Hi, if I understand correctly, once we have the edges, we already have the segmentation. So may I ask what you mean by the last sentence, do we need to refine the segments from the edges?", "273004": "What does FWIW means?", "273347": "Nevermind I was able to figure out what FWIW is. Thanks.", "273640": "Yan Wang, In my case at least, I'm predicting the edges as a bit mask. So I end up with one, approximate, bit mask of the edges. And another of the foreground. Just having the bitmap of the edges doesn't automatically give me the masks for each nuclei, I still have to split them up. And that is the fiddly part.", "273641": "WilliamGreen, glad you figured it out. It's one of those very technical abbreviations ;-)", "274656": "Hi, Tim Hochberg, I get the boundary through mask minus erosion in 1pixel or then dilate them in 2-3 pixels, but I try many deep models it always hardly to find the boundary..Do u have any idea?", "274850": "bonsen, I'm actually using between 4 and 10 pixel wide edges (still playing around), so perhaps just try using larger edges. Also when I'm selecting crops to feed to the net at training time, I arrange to give it crops that contain an edge at least half the time.", "276585": "For detectron you need a GPU with RAM 4GB or more. I just went thru installing it thru docker and with my 2GB GPU, I couldn't even able to run inference demo. Probably for training you may need multiple GPUs.", "277020": "check this latest paper!\nhttps://arxiv.org/pdf/1802.00383.pdf\n\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/277020/8471/2.png\n [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/277020/8470/1.png", "277087": "When u get the seed and the segmentation of forgorund, how to mix them for different colors?", "277088": "I will try, thanks.", "277169": "Hi, Tim Hochberg, \nI'm sorry that I come to bother u again..... \nIf I use wide edges to train and I will get the wide edges after prediction. It seems that I can't use it to guide the segmentation.\nThanks", "277596": "I don't think the stated results in paper is easy to reproduce. Feel too good to be true.", "277761": "Heng CherKeng: What is your best result using mask rcnn? Thanks", "277765": "i haven't made a mask-rcnn submission yet. \n\ni am fine-tunning my code, and now only doing training/validation on gray images (and not the color ones)\n\non local validation set, i have score of about 0.65. (from previous results, this validation set is about 0.1 higher than LB score)", "277804": "Great. I think your score will be better when you split data into some clusters (3 to 5) as someone did. However, it must create 3/5 networks which used for each cluster", "277808": "i just posted some visual results at my thread:\n\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/47686\n\nsee \"i try to write a pytorch mask-rcnn from scratch. ....\nHere are some results from rpn(red is before nms, white is after) and rcnn (pink).\"", "277853": "when u extract the object pixels . should i try to seperate them to many images like in the dataset, is that neceesary?", "278569": "Hi Ertuka. Do you use Detectron for this dataset. It must be convert to COCO format. Do you have any idea to do it? Thanks", "279496": "john1231983 I have failed to be able to run Detectron demo on my GPU since requires at least 4GB RAM so never tried it further. Possibly one may more than 6GB for any chance to apply to this", "279509": "Hi @Heng, This article looks very promising. Have you tried it?", "279690": "I generated the mask inside objects and contours with unet. But I don't know how to use the contour to split different objects. Can you describe your edge guided segmentation method? Thanks.", "279721": "zmingen, I use something pretty similar to what's described here: http://scikit-image.org/docs/dev/auto_examples/segmentation/plot_watershed.html. I use the background - edges to generate seeds, and then to watershed segmentation to fill out from the seeds.", "284980": "this paper has better results than mask rcnn:\n\nMaskLab: Instance Segmentation by Refining Object Detection with Semantic and Direction Features\n\nhttp://liangchiehchen.com/", "286308": "but they have not released the code yet!", "1370634": "Hey! Really faultless Mask RCNN installation for TensorFlow GPU\n 👉 https://github.com/BuseYaren/Installation-MaskRCNN" }, "source": "meta" }![\"enter](\"http://img.blog.csdn.net/20170702200353295?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYmVhX3RyZWU=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\")
![\"enter](\"http://img.blog.csdn.net/20170703222133261?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYmVhX3RyZWU=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\")
![\"enter](\"http://img.blog.csdn.net/20170703223354397?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYmVhX3RyZWU=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\")
![\"enter](\"http://img.blog.csdn.net/20171016102609952?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\")
![\"enter](\"http://img.blog.csdn.net/20171016101725562?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\")
![\"enter](\"http://img.blog.csdn.net/20171016101647474?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast\")