The third place solution, tie with #2 jacobkie achieving 0.614 on the Private Leader-board, is based on a single Mask-RCNN model using as code-base Matterport's Mask-RCNN (https://github.com/matterport/Mask_RCNN).
\n\nSummary
\n\nI don't think I have done many different things to what others have reported and my solution is quite simple. I'll be happy to share the code once I have the approval by kaggle and sponsors. I tried many different things, but the main two contributions are:
\n\n1) Strong scaling augmentation, a lot of zooming in and out and aspect ratio changes before taking the 512x512 crops used as inputs to the model during training.
\n\n2) Test time augmentation, I used 15 different augmentations at test time with different rotations, scalings, channel color shifts, etc. This takes a loooong time (aprox. 2 days for the stage_2 test set) and a binary dilation post-processing actually gives a very similar score, so I would use the latter if asked now (although it is easy to tell now that we can see the PL scores..)
\n\nTraining data
\n\nNo external data was used, only stage 1 training set. I used the corrected data-set compiled in: https://github.com/lopuhin/kaggle-dsbowl-2018-dataset-fixes \nI didn't have the time to include any of the external data, I was also afraid that the different annotation styles might introduced unwanted bias in the predictions.\nNo prepocessing.
\n\nAugmentations
\n\nIn addition to the scaling augmentation mentioned above I used left-right and up-down flips, random 90 degree rotations, random additional rotation on top of those, random channel color shifts
\n\nParameters
\n\nHere are some of the parameters for comparison (https://www.kaggle.com/c/data-science-bowl-2018/discussion/54920):
\n\nCodeBase Type-1 and 2\nMEAN_PIXEL [123.7, 116,8, 103,9]\nLEARNING_RATE Start 0.001 and down to 3*10^-5\nLEARNING_SCHEDULE ~120 always \"all\"\nRPN_ANCHOR_RATIOS [0.5, 1, 2]\nUSE_MINI_MASK True\nMINI_MASK_SHAPE (56,56)\nGPU_COUNT 1\nIMAGES_PER_GPU 2\nSTEPS_PER_EPOCH 332\nVALIDATION_STEPS 0\nBACKBONE resnet101\nNUM_CLASSES 1+1\nIMAGE_MIN_DIM 512\nIMAGE_MAX_DIM Not used\nIMAGE_PADDING Not used\nRPN_ANCHOR_SCALES 8,16,32,64,128\nRPN_ANCHOR_STRIDE 1\nBACKBONE_STRIDES 4,8,16,32,64\nRPN_TRAIN_ANCHORS_PER_IMAGE 256\nIMAGE_MIN_SCALE Not used\nIMAGE_RESIZE_MODE crop at training, pad64 for inference\nRPN_NMS_THRESHOLD 0.7\nDETECTION_MIN_CONFIDENCE 0.9\nDETECTION_NMS_THRESHOLD 0.2\nTRAIN_ROIS_PER_IMAGE 600\nDETECTION_MAX_INSTANCES 512\nMAX_GT_INSTANCES 256\ninit_with coco\nDATA_AUGMENTATION scaling, crop, flip-lr, flip-up, 90 rotation, rotation, channel_shift
\n\nSource code is now available follow this link to github
", "messages": [ { "id": 326131, "postDate": "2018-05-09T10:31:53.703Z", "content": "The third place solution, tie with #2 jacobkie achieving 0.614 on the Private Leader-board, is based on a single Mask-RCNN model using as code-base Matterport's Mask-RCNN (https://github.com/matterport/Mask_RCNN).
\n\nSummary
\n\nI don't think I have done many different things to what others have reported and my solution is quite simple. I'll be happy to share the code once I have the approval by kaggle and sponsors. I tried many different things, but the main two contributions are:
\n\n1) Strong scaling augmentation, a lot of zooming in and out and aspect ratio changes before taking the 512x512 crops used as inputs to the model during training.
\n\n2) Test time augmentation, I used 15 different augmentations at test time with different rotations, scalings, channel color shifts, etc. This takes a loooong time (aprox. 2 days for the stage_2 test set) and a binary dilation post-processing actually gives a very similar score, so I would use the latter if asked now (although it is easy to tell now that we can see the PL scores..)
\n\nTraining data
\n\nNo external data was used, only stage 1 training set. I used the corrected data-set compiled in: https://github.com/lopuhin/kaggle-dsbowl-2018-dataset-fixes \nI didn't have the time to include any of the external data, I was also afraid that the different annotation styles might introduced unwanted bias in the predictions.\nNo prepocessing.
\n\nAugmentations
\n\nIn addition to the scaling augmentation mentioned above I used left-right and up-down flips, random 90 degree rotations, random additional rotation on top of those, random channel color shifts
\n\nParameters
\n\nHere are some of the parameters for comparison (https://www.kaggle.com/c/data-science-bowl-2018/discussion/54920):
\n\nCodeBase Type-1 and 2\nMEAN_PIXEL [123.7, 116,8, 103,9]\nLEARNING_RATE Start 0.001 and down to 3*10^-5\nLEARNING_SCHEDULE ~120 always \"all\"\nRPN_ANCHOR_RATIOS [0.5, 1, 2]\nUSE_MINI_MASK True\nMINI_MASK_SHAPE (56,56)\nGPU_COUNT 1\nIMAGES_PER_GPU 2\nSTEPS_PER_EPOCH 332\nVALIDATION_STEPS 0\nBACKBONE resnet101\nNUM_CLASSES 1+1\nIMAGE_MIN_DIM 512\nIMAGE_MAX_DIM Not used\nIMAGE_PADDING Not used\nRPN_ANCHOR_SCALES 8,16,32,64,128\nRPN_ANCHOR_STRIDE 1\nBACKBONE_STRIDES 4,8,16,32,64\nRPN_TRAIN_ANCHORS_PER_IMAGE 256\nIMAGE_MIN_SCALE Not used\nIMAGE_RESIZE_MODE crop at training, pad64 for inference\nRPN_NMS_THRESHOLD 0.7\nDETECTION_MIN_CONFIDENCE 0.9\nDETECTION_NMS_THRESHOLD 0.2\nTRAIN_ROIS_PER_IMAGE 600\nDETECTION_MAX_INSTANCES 512\nMAX_GT_INSTANCES 256\ninit_with coco\nDATA_AUGMENTATION scaling, crop, flip-lr, flip-up, 90 rotation, rotation, channel_shift
\n\nSource code is now available follow this link to github
", "rawMarkdown": "The third place solution, tie with #2 jacobkie achieving 0.614 on the Private Leader-board, is based on a single Mask-RCNN model using as code-base Matterport's Mask-RCNN (https://github.com/matterport/Mask_RCNN). \n\n**Summary**\n\n\n----------\n\n\nI don't think I have done many different things to what others have reported and my solution is quite simple. I'll be happy to share the code once I have the approval by kaggle and sponsors. I tried many different things, but the main two contributions are:\n\n1) **Strong scaling augmentation**, a lot of zooming in and out and aspect ratio changes before taking the 512x512 crops used as inputs to the model during training.\n\n2) **Test time augmentation**, I used 15 different augmentations at test time with different rotations, scalings, channel color shifts, etc. This takes a loooong time (aprox. 2 days for the stage_2 test set) and a binary dilation post-processing actually gives a very similar score, so I would use the latter if asked now (although it is easy to tell now that we can see the PL scores..)\n\n\n**Training data**\n\n\n----------\n\n\nNo external data was used, only stage 1 training set. I used the corrected data-set compiled in: https://github.com/lopuhin/kaggle-dsbowl-2018-dataset-fixes \nI didn't have the time to include any of the external data, I was also afraid that the different annotation styles might introduced unwanted bias in the predictions.\nNo prepocessing.\n\n**Augmentations**\n\n\n----------\n\n\nIn addition to the scaling augmentation mentioned above I used left-right and up-down flips, random 90 degree rotations, random additional rotation on top of those, random channel color shifts \n\n**Parameters**\n\n\n----------\n\n\nHere are some of the parameters for comparison (https://www.kaggle.com/c/data-science-bowl-2018/discussion/54920):\n\nCodeBase\tType-1 and 2\nMEAN_PIXEL\t[123.7, 116,8, 103,9]\nLEARNING_RATE\tStart 0.001 and down to 3*10^-5\nLEARNING_SCHEDULE\t~120 always \"all\"\nRPN_ANCHOR_RATIOS\t[0.5, 1, 2]\nUSE_MINI_MASK\tTrue\nMINI_MASK_SHAPE\t(56,56)\nGPU_COUNT\t1\nIMAGES_PER_GPU\t2\nSTEPS_PER_EPOCH\t332\nVALIDATION_STEPS\t0\nBACKBONE\tresnet101\nNUM_CLASSES\t1+1\nIMAGE_MIN_DIM\t512\nIMAGE_MAX_DIM\tNot used\nIMAGE_PADDING\tNot used\nRPN_ANCHOR_SCALES\t8,16,32,64,128\nRPN_ANCHOR_STRIDE\t1\nBACKBONE_STRIDES\t4,8,16,32,64\nRPN_TRAIN_ANCHORS_PER_IMAGE\t256\nIMAGE_MIN_SCALE\tNot used\nIMAGE_RESIZE_MODE\tcrop at training, pad64 for inference\nRPN_NMS_THRESHOLD\t0.7\nDETECTION_MIN_CONFIDENCE\t0.9\nDETECTION_NMS_THRESHOLD\t0.2\nTRAIN_ROIS_PER_IMAGE\t600\nDETECTION_MAX_INSTANCES\t512\nMAX_GT_INSTANCES\t256\ninit_with\tcoco\nDATA_AUGMENTATION\tscaling, crop, flip-lr, flip-up, 90 rotation, rotation, channel_shift\n\n----------\nSource code is now available [follow this link to github][1]\n\n\n\n [1]: https://github.com/Gelu74/DSB_2018", "votes": 36 }, { "id": 328630, "postDate": "2018-05-14T18:48:29.877Z", "content": "Congrats! Regarding test time augmentation, how did you ensemble the predictions from the different settings?
", "rawMarkdown": "Congrats! Regarding test time augmentation, how did you ensemble the predictions from the different settings?", "votes": 3, "replies": [ { "id": 346917, "postDate": "2018-06-22T18:18:58.710Z", "content": "I second this comment. Ensembling the output of matterport mask rcnn is certainly not a trivial task. Any hints would be great :)
", "rawMarkdown": "I second this comment. Ensembling the output of matterport mask rcnn is certainly not a trivial task. Any hints would be great :)\n" }, { "id": 358526, "postDate": "2018-07-18T10:22:00.297Z", "content": "I have updated the description with a link to the source code where you can view the test time augmentation ensembling. As I mention in the README, this test time augmentation is quite tricky in that it has quite a few parameters to tune and a simple binary dilation operation performs better, although it is easy to say now that we have access to the private leaderboard scores..
", "rawMarkdown": "I have updated the description with a link to the source code where you can view the test time augmentation ensembling. As I mention in the README, this test time augmentation is quite tricky in that it has quite a few parameters to tune and a simple binary dilation operation performs better, although it is easy to say now that we have access to the private leaderboard scores.." }, { "id": 540344, "postDate": "2019-05-31T11:30:12.957Z", "content": "I'm so stupid to understand how to ensemble the tta prediction.
", "rawMarkdown": "I'm so stupid to understand how to ensemble the tta prediction." } ] }, { "id": 326245, "postDate": "2018-05-09T13:30:21.197Z", "content": "Nice work and congrats! Great to see exactly what parms you used.
", "rawMarkdown": "Nice work and congrats! Great to see exactly what parms you used.", "votes": 3 }, { "id": 328494, "postDate": "2018-05-14T13:25:55.623Z", "content": "I love the simplicity of the solution. Thanks!
", "rawMarkdown": "I love the simplicity of the solution. Thanks!", "votes": 1 }, { "id": 326425, "postDate": "2018-05-09T17:31:53.680Z", "content": "@Deep Retina Congratulations again and thank you very much for sharing your parameters! I will add them the page:
\n\n\n\nOne quick question: we saw most contestants were using one of these two versions of Matterport Mask RCNN.
\n\nType 1: Older Mask_RCNN_code from here: https://github.com/killthekitten/kaggle-ds-bowl-2018-baseline
\n\nType 2: latest Mask_RCNN_code from here: https://github.com/matterport/Mask_RCNN
\n\nType 2: has the latest updates from Waleed while type1 was a copy of the code from Waleed's original implementation.\nType 1 was not a fork; so i am not sure at what point exactly it was created.
\n\nCould you please say which one you used ?
\n\nThanks again ! \nMehul
", "rawMarkdown": "@Deep Retina Congratulations again and thank you very much for sharing your parameters! I will add them the page: \n\nhttps://github.com/mpsampat/DSB_2018_Parameter_Selection_For_Matterport_Mask_RCNN/blob/master/comparison_of_Matterport_Mask_RCNN_parameters.md\n\nOne quick question: we saw most contestants were using one of these two versions of Matterport Mask RCNN. \n\nType 1: Older Mask_RCNN_code from here: https://github.com/killthekitten/kaggle-ds-bowl-2018-baseline\n\nType 2: latest Mask_RCNN_code from here: https://github.com/matterport/Mask_RCNN\n\nType 2: has the latest updates from Waleed while type1 was a copy of the code from Waleed's original implementation.\nType 1 was not a fork; so i am not sure at what point exactly it was created.\n\nCould you please say which one you used ? \n\nThanks again ! \nMehul", "votes": 1, "replies": [ { "id": 326435, "postDate": "2018-05-09T17:43:25.650Z", "content": "I actually used a bit of both... I started using the code from https://github.com/matterport/Mask_RCNN at the beginning of the competition and worked on a clone of that repository on my computer for most of the competition (which I guess is close to Type 1), but that version lacked the \"pad64\" for predicting on variable size inputs so a few days before the competition ended I moved to the latest version. I have organized all my code and I now have a working version which is a Type 2 with a few changes in the code. I hope to share that code during next week.
", "rawMarkdown": "I actually used a bit of both... I started using the code from https://github.com/matterport/Mask_RCNN at the beginning of the competition and worked on a clone of that repository on my computer for most of the competition (which I guess is close to Type 1), but that version lacked the \"pad64\" for predicting on variable size inputs so a few days before the competition ended I moved to the latest version. I have organized all my code and I now have a working version which is a Type 2 with a few changes in the code. I hope to share that code during next week.", "votes": 5 } ] }, { "id": 374818, "postDate": "2018-08-23T23:06:34.853Z", "content": "Congratulations @Deep Retina team, and thanks for sharing your solution and code.
", "rawMarkdown": "Congratulations @Deep Retina team, and thanks for sharing your solution and code." }, { "id": 359337, "postDate": "2018-07-19T22:50:57.090Z", "content": "hi, Where can i find weights for this submission? I have similar objects to detect and wondering if I can reuse the weights to inference my pictures.
", "rawMarkdown": "hi, Where can i find weights for this submission? I have similar objects to detect and wondering if I can reuse the weights to inference my pictures. ", "replies": [ { "id": 359540, "postDate": "2018-07-20T10:04:41.863Z", "content": "I have modified the README with a link to the weights, the file is too large to upload to github
", "rawMarkdown": "I have modified the README with a link to the weights, the file is too large to upload to github" }, { "id": 365451, "postDate": "2018-08-02T16:38:07.523Z", "content": "hi, thanks. but where is the readme file located? i didn't find the link to weights in https://github.com/killthekitten/kaggle-ds-bowl-2018-baseline. or https://github.com/matterport/Mask_RCNN. Could you please paste the link here if its possible. Appreciate it.
", "rawMarkdown": "hi, thanks. but where is the readme file located? i didn't find the link to weights in https://github.com/killthekitten/kaggle-ds-bowl-2018-baseline. or https://github.com/matterport/Mask_RCNN. Could you please paste the link here if its possible. Appreciate it. " } ] }, { "id": 330283, "postDate": "2018-05-18T13:28:55.120Z", "content": "Fantastic result and work, even more so it being a single model solution! Congrats!
", "rawMarkdown": "Fantastic result and work, even more so it being a single model solution! Congrats!" }, { "id": 394271, "postDate": "2018-09-26T15:27:59.243Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 328630, "author_name": "Naoto Usuyama", "author_url": "", "post_date": "2018-05-14T18:48:29.877000", "content": "Congrats! Regarding test time augmentation, how did you ensemble the predictions from the different settings?
", "votes": 3, "replies": [ { "id": 346917, "author_name": "SimonGraham", "author_url": "", "post_date": "2018-06-22T18:18:58.710000", "content": "I second this comment. Ensembling the output of matterport mask rcnn is certainly not a trivial task. Any hints would be great :)
", "votes": 0, "replies": [] }, { "id": 358526, "author_name": "Deep Retina", "author_url": "", "post_date": "2018-07-18T10:22:00.297000", "content": "I have updated the description with a link to the source code where you can view the test time augmentation ensembling. As I mention in the README, this test time augmentation is quite tricky in that it has quite a few parameters to tune and a simple binary dilation operation performs better, although it is easy to say now that we have access to the private leaderboard scores..
", "votes": 0, "replies": [] }, { "id": 540344, "author_name": "xiatian", "author_url": "", "post_date": "2019-05-31T11:30:12.957000", "content": "I'm so stupid to understand how to ensemble the tta prediction.
", "votes": 0, "replies": [] } ] }, { "id": 326245, "author_name": "Russ Wolfinger", "author_url": "", "post_date": "2018-05-09T13:30:21.197000", "content": "Nice work and congrats! Great to see exactly what parms you used.
", "votes": 3, "replies": [] }, { "id": 328494, "author_name": "Andrés Miguel Torrubia Sáez", "author_url": "", "post_date": "2018-05-14T13:25:55.623000", "content": "I love the simplicity of the solution. Thanks!
", "votes": 1, "replies": [] }, { "id": 326425, "author_name": "Mehul Sampat", "author_url": "", "post_date": "2018-05-09T17:31:53.680000", "content": "@Deep Retina Congratulations again and thank you very much for sharing your parameters! I will add them the page:
\n\n\n\nOne quick question: we saw most contestants were using one of these two versions of Matterport Mask RCNN.
\n\nType 1: Older Mask_RCNN_code from here: https://github.com/killthekitten/kaggle-ds-bowl-2018-baseline
\n\nType 2: latest Mask_RCNN_code from here: https://github.com/matterport/Mask_RCNN
\n\nType 2: has the latest updates from Waleed while type1 was a copy of the code from Waleed's original implementation.\nType 1 was not a fork; so i am not sure at what point exactly it was created.
\n\nCould you please say which one you used ?
\n\nThanks again ! \nMehul
", "votes": 1, "replies": [ { "id": 326435, "author_name": "Deep Retina", "author_url": "", "post_date": "2018-05-09T17:43:25.650000", "content": "I actually used a bit of both... I started using the code from https://github.com/matterport/Mask_RCNN at the beginning of the competition and worked on a clone of that repository on my computer for most of the competition (which I guess is close to Type 1), but that version lacked the \"pad64\" for predicting on variable size inputs so a few days before the competition ended I moved to the latest version. I have organized all my code and I now have a working version which is a Type 2 with a few changes in the code. I hope to share that code during next week.
", "votes": 5, "replies": [] } ] }, { "id": 374818, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2018-08-23T23:06:34.853000", "content": "Congratulations @Deep Retina team, and thanks for sharing your solution and code.
", "votes": 0, "replies": [] }, { "id": 359337, "author_name": "BearlyAI", "author_url": "", "post_date": "2018-07-19T22:50:57.090000", "content": "hi, Where can i find weights for this submission? I have similar objects to detect and wondering if I can reuse the weights to inference my pictures.
", "votes": 0, "replies": [ { "id": 359540, "author_name": "Deep Retina", "author_url": "", "post_date": "2018-07-20T10:04:41.863000", "content": "I have modified the README with a link to the weights, the file is too large to upload to github
", "votes": 0, "replies": [] }, { "id": 365451, "author_name": "BearlyAI", "author_url": "", "post_date": "2018-08-02T16:38:07.523000", "content": "hi, thanks. but where is the readme file located? i didn't find the link to weights in https://github.com/killthekitten/kaggle-ds-bowl-2018-baseline. or https://github.com/matterport/Mask_RCNN. Could you please paste the link here if its possible. Appreciate it.
", "votes": 0, "replies": [] } ] }, { "id": 330283, "author_name": "Jack Vial", "author_url": "", "post_date": "2018-05-18T13:28:55.120000", "content": "Fantastic result and work, even more so it being a single model solution! Congrats!
", "votes": 0, "replies": [] }, { "id": 394271, "author_name": "", "author_url": "", "post_date": "2018-09-26T15:27:59.243000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "326131": "The third place solution, tie with #2 jacobkie achieving 0.614 on the Private Leader-board, is based on a single Mask-RCNN model using as code-base Matterport's Mask-RCNN (https://github.com/matterport/Mask_RCNN). \n\n**Summary**\n\n\n----------\n\n\nI don't think I have done many different things to what others have reported and my solution is quite simple. I'll be happy to share the code once I have the approval by kaggle and sponsors. I tried many different things, but the main two contributions are:\n\n1) **Strong scaling augmentation**, a lot of zooming in and out and aspect ratio changes before taking the 512x512 crops used as inputs to the model during training.\n\n2) **Test time augmentation**, I used 15 different augmentations at test time with different rotations, scalings, channel color shifts, etc. This takes a loooong time (aprox. 2 days for the stage_2 test set) and a binary dilation post-processing actually gives a very similar score, so I would use the latter if asked now (although it is easy to tell now that we can see the PL scores..)\n\n\n**Training data**\n\n\n----------\n\n\nNo external data was used, only stage 1 training set. I used the corrected data-set compiled in: https://github.com/lopuhin/kaggle-dsbowl-2018-dataset-fixes \nI didn't have the time to include any of the external data, I was also afraid that the different annotation styles might introduced unwanted bias in the predictions.\nNo prepocessing.\n\n**Augmentations**\n\n\n----------\n\n\nIn addition to the scaling augmentation mentioned above I used left-right and up-down flips, random 90 degree rotations, random additional rotation on top of those, random channel color shifts \n\n**Parameters**\n\n\n----------\n\n\nHere are some of the parameters for comparison (https://www.kaggle.com/c/data-science-bowl-2018/discussion/54920):\n\nCodeBase\tType-1 and 2\nMEAN_PIXEL\t[123.7, 116,8, 103,9]\nLEARNING_RATE\tStart 0.001 and down to 3*10^-5\nLEARNING_SCHEDULE\t~120 always \"all\"\nRPN_ANCHOR_RATIOS\t[0.5, 1, 2]\nUSE_MINI_MASK\tTrue\nMINI_MASK_SHAPE\t(56,56)\nGPU_COUNT\t1\nIMAGES_PER_GPU\t2\nSTEPS_PER_EPOCH\t332\nVALIDATION_STEPS\t0\nBACKBONE\tresnet101\nNUM_CLASSES\t1+1\nIMAGE_MIN_DIM\t512\nIMAGE_MAX_DIM\tNot used\nIMAGE_PADDING\tNot used\nRPN_ANCHOR_SCALES\t8,16,32,64,128\nRPN_ANCHOR_STRIDE\t1\nBACKBONE_STRIDES\t4,8,16,32,64\nRPN_TRAIN_ANCHORS_PER_IMAGE\t256\nIMAGE_MIN_SCALE\tNot used\nIMAGE_RESIZE_MODE\tcrop at training, pad64 for inference\nRPN_NMS_THRESHOLD\t0.7\nDETECTION_MIN_CONFIDENCE\t0.9\nDETECTION_NMS_THRESHOLD\t0.2\nTRAIN_ROIS_PER_IMAGE\t600\nDETECTION_MAX_INSTANCES\t512\nMAX_GT_INSTANCES\t256\ninit_with\tcoco\nDATA_AUGMENTATION\tscaling, crop, flip-lr, flip-up, 90 rotation, rotation, channel_shift\n\n----------\nSource code is now available [follow this link to github][1]\n\n\n\n [1]: https://github.com/Gelu74/DSB_2018", "328630": "Congrats! Regarding test time augmentation, how did you ensemble the predictions from the different settings?", "326245": "Nice work and congrats! Great to see exactly what parms you used.", "328494": "I love the simplicity of the solution. Thanks!", "326425": "@Deep Retina Congratulations again and thank you very much for sharing your parameters! I will add them the page: \n\nhttps://github.com/mpsampat/DSB_2018_Parameter_Selection_For_Matterport_Mask_RCNN/blob/master/comparison_of_Matterport_Mask_RCNN_parameters.md\n\nOne quick question: we saw most contestants were using one of these two versions of Matterport Mask RCNN. \n\nType 1: Older Mask_RCNN_code from here: https://github.com/killthekitten/kaggle-ds-bowl-2018-baseline\n\nType 2: latest Mask_RCNN_code from here: https://github.com/matterport/Mask_RCNN\n\nType 2: has the latest updates from Waleed while type1 was a copy of the code from Waleed's original implementation.\nType 1 was not a fork; so i am not sure at what point exactly it was created.\n\nCould you please say which one you used ? \n\nThanks again ! \nMehul", "374818": "Congratulations @Deep Retina team, and thanks for sharing your solution and code.", "359337": "hi, Where can i find weights for this submission? I have similar objects to detect and wondering if I can reuse the weights to inference my pictures. ", "330283": "Fantastic result and work, even more so it being a single model solution! Congrats!", "394271": "" }, "complete": true, "source": "existing_discussion", "competition_id": 8089, "competition_slug": "data-science-bowl-2018", "team_id": 1265104, "solution_url": "/competitions/data-science-bowl-2018/writeups/deep-retina-3th-place-on-private-lb-matterport-s-m" }