This competition is a tough one, with more than 4K resolution images training a decent model will take ages. One look at the competition and the immediate thought of solution is directly Mask RCNN based on the multi-class segmentation problem.
\n\nOur solution is a simple RESNET50 and RESNET101 Mask RCNN ensemble, training didn't stop improving for both of the models but we were forced to stop halfway after a few epochs wasting a lot of time as predicting and generating predictions also takes a lot of time. The ensemble we used is a simple average of mask and weighted average of scores based on proximity.
\n\nThe ensemble didn't help as much as we expected it to and we only had time to test one type of parameters. Looks like we should have used our time to train more epochs. All in all a great experience especially with the lack of code from kernals to go forward. Looking forward to the Top Teams Sharing!
", "messages": [ { "id": "341943", "postDate": "06/12/2018 14:54:31", "content": "This competition is a tough one, with more than 4K resolution images training a decent model will take ages. One look at the competition and the immediate thought of solution is directly Mask RCNN based on the multi-class segmentation problem.
\n\nOur solution is a simple RESNET50 and RESNET101 Mask RCNN ensemble, training didn't stop improving for both of the models but we were forced to stop halfway after a few epochs wasting a lot of time as predicting and generating predictions also takes a lot of time. The ensemble we used is a simple average of mask and weighted average of scores based on proximity.
\n\nThe ensemble didn't help as much as we expected it to and we only had time to test one type of parameters. Looks like we should have used our time to train more epochs. All in all a great experience especially with the lack of code from kernals to go forward. Looking forward to the Top Teams Sharing!
", "rawMarkdown": "This competition is a tough one, with more than 4K resolution images training a decent model will take ages. One look at the competition and the immediate thought of solution is directly Mask RCNN based on the multi-class segmentation problem. \n\nOur solution is a simple RESNET50 and RESNET101 Mask RCNN ensemble, training didn't stop improving for both of the models but we were forced to stop halfway after a few epochs wasting a lot of time as predicting and generating predictions also takes a lot of time. The ensemble we used is a simple average of mask and weighted average of scores based on proximity. \n\nThe ensemble didn't help as much as we expected it to and we only had time to test one type of parameters. Looks like we should have used our time to train more epochs. All in all a great experience especially with the lack of code from kernals to go forward. Looking forward to the Top Teams Sharing!", "votes": null }, { "id": "342473", "postDate": "06/13/2018 15:07:24", "content": "Yeah, definitely a tough challenge as good part and the bad part of this challenge is it doesn't stand for ranking point.
\n\nWell, my model is composed of simple UNET + Mask-RCNN and yes, ensembling didn't work for this challenge anyway. But I would love to see solutions of top scorers.
", "rawMarkdown": "Yeah, definitely a tough challenge as good part and the bad part of this challenge is it doesn't stand for ranking point.\n\nWell, my model is composed of simple UNET + Mask-RCNN and yes, ensembling didn't work for this challenge anyway. But I would love to see solutions of top scorers.", "votes": null }, { "id": "342493", "postDate": "06/13/2018 15:40:02", "content": "esemble can work
", "rawMarkdown": "esemble can work", "votes": null }, { "id": "342503", "postDate": "06/13/2018 15:55:28", "content": "Ensembling worked for me just not that much . @Tommy Would love to hear your solution 🙂
", "rawMarkdown": "Ensembling worked for me just not that much . @Tommy Would love to hear your solution 🙂", "votes": null }, { "id": "342512", "postDate": "06/13/2018 16:09:00", "content": "I didn't tried ensembling cus of submission problem during the last day of challenge.
", "rawMarkdown": "I didn't tried ensembling cus of submission problem during the last day of challenge.", "votes": null }, { "id": "342667", "postDate": "06/13/2018 21:33:34", "content": "Ensemble works for me too
", "rawMarkdown": "Ensemble works for me too", "votes": null } ], "comments": [ { "id": 342473, "author_name": "amitkumarjaiswal", "author_url": "", "post_date": "06/13/2018 15:07:24", "content": "Yeah, definitely a tough challenge as good part and the bad part of this challenge is it doesn't stand for ranking point.
\n\nWell, my model is composed of simple UNET + Mask-RCNN and yes, ensembling didn't work for this challenge anyway. But I would love to see solutions of top scorers.
", "votes": null, "replies": [ { "id": 342493, "author_name": "zhuangyq", "author_url": "", "post_date": "06/13/2018 15:40:02", "content": "esemble can work
", "votes": null, "replies": [] }, { "id": 342503, "author_name": "dicksonchin93", "author_url": "", "post_date": "06/13/2018 15:55:28", "content": "Ensembling worked for me just not that much . @Tommy Would love to hear your solution 🙂
", "votes": null, "replies": [] }, { "id": 342512, "author_name": "amitkumarjaiswal", "author_url": "", "post_date": "06/13/2018 16:09:00", "content": "I didn't tried ensembling cus of submission problem during the last day of challenge.
", "votes": null, "replies": [] }, { "id": 342667, "author_name": "woodywang", "author_url": "", "post_date": "06/13/2018 21:33:34", "content": "Ensemble works for me too
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "341943": "This competition is a tough one, with more than 4K resolution images training a decent model will take ages. One look at the competition and the immediate thought of solution is directly Mask RCNN based on the multi-class segmentation problem. \n\nOur solution is a simple RESNET50 and RESNET101 Mask RCNN ensemble, training didn't stop improving for both of the models but we were forced to stop halfway after a few epochs wasting a lot of time as predicting and generating predictions also takes a lot of time. The ensemble we used is a simple average of mask and weighted average of scores based on proximity. \n\nThe ensemble didn't help as much as we expected it to and we only had time to test one type of parameters. Looks like we should have used our time to train more epochs. All in all a great experience especially with the lack of code from kernals to go forward. Looking forward to the Top Teams Sharing!", "342473": "Yeah, definitely a tough challenge as good part and the bad part of this challenge is it doesn't stand for ranking point.\n\nWell, my model is composed of simple UNET + Mask-RCNN and yes, ensembling didn't work for this challenge anyway. But I would love to see solutions of top scorers.", "342493": "esemble can work", "342503": "Ensembling worked for me just not that much . @Tommy Would love to hear your solution 🙂", "342512": "I didn't tried ensembling cus of submission problem during the last day of challenge.", "342667": "Ensemble works for me too" }, "source": "meta" }