So I have approached the problem using fastrcnn resnet101 backbone (since it was given better results than resnet50) and have applied different augmentation techniques with the WBF ensembling.\nStill, my score cant cross the 0.7 mark.\nSo I was thinking how much can you score just by using fastrcnn.\nI don't know why my score has got stagnant in a single value.
", "messages": [ { "id": "891976", "postDate": "06/18/2020 15:11:08", "content": "So I have approached the problem using fastrcnn resnet101 backbone (since it was given better results than resnet50) and have applied different augmentation techniques with the WBF ensembling.\nStill, my score cant cross the 0.7 mark.\nSo I was thinking how much can you score just by using fastrcnn.\nI don't know why my score has got stagnant in a single value.
", "rawMarkdown": "So I have approached the problem using fastrcnn resnet101 backbone (since it was given better results than resnet50) and have applied different augmentation techniques with the **WBF** ensembling.\nStill, my score cant cross the 0.7 mark.\nSo I was thinking how much can you score just by using fastrcnn.\nI don't know why my score has got stagnant in a single value.", "votes": null }, { "id": "892100", "postDate": "06/18/2020 16:47:56", "content": "What image resolution are you training?
", "rawMarkdown": "What image resolution are you training?", "votes": null }, { "id": "892150", "postDate": "06/18/2020 17:19:48", "content": "on full image
", "rawMarkdown": "on full image", "votes": null }, { "id": "892227", "postDate": "06/18/2020 18:12:16", "content": "I’m at over 0.7 using fasterrcnn with resnet50. Single model, no TTA, no pseudolabeling.
\n\nJust have to keep experimenting
", "rawMarkdown": "I’m at over 0.7 using fasterrcnn with resnet50. Single model, no TTA, no pseudolabeling.\n\nJust have to keep experimenting", "votes": null }, { "id": "892247", "postDate": "06/18/2020 18:31:27", "content": "Have you used any ensembling techniques like wbf?
", "rawMarkdown": "Have you used any ensembling techniques like wbf?", "votes": null }, { "id": "892320", "postDate": "06/18/2020 19:57:17", "content": "Yes, I've been using this since before it was made \"public\". Though that's not included in the >0.7 score I mentioned.
", "rawMarkdown": "Yes, I've been using this since before it was made \"public\". Though that's not included in the >0.7 score I mentioned.", "votes": null }, { "id": "892556", "postDate": "06/19/2020 02:15:11", "content": "mine: 0.7314 no TTA, pseudo, wbf
", "rawMarkdown": "mine: 0.7314 no TTA, pseudo, wbf", "votes": null }, { "id": "892681", "postDate": "06/19/2020 04:38:51", "content": "Impressive! I take it you're back in? 😉
", "rawMarkdown": "Impressive! I take it you're back in? 😉", "votes": null }, { "id": "892771", "postDate": "06/19/2020 06:28:12", "content": "Hopefully I won't give up this time. :))
", "rawMarkdown": "Hopefully I won't give up this time. :))", "votes": null }, { "id": "892870", "postDate": "06/19/2020 08:14:54", "content": "that's really impressive\nI will keep trying
", "rawMarkdown": "that's really impressive\nI will keep trying", "votes": null }, { "id": "893420", "postDate": "06/19/2020 15:33:42", "content": "when I removed the boxes ( not that image) with abnormal sizes my lb score reduced., not getting why😩
", "rawMarkdown": "when I removed the boxes ( not that image) with abnormal sizes my lb score reduced., not getting why😩", "votes": null }, { "id": "893485", "postDate": "06/19/2020 16:40:04", "content": "What do you mean you remove boxes with abnormal sizes? From the training data? From your predictions?
", "rawMarkdown": "What do you mean you remove boxes with abnormal sizes? From the training data? From your predictions?", "votes": null }, { "id": "893524", "postDate": "06/19/2020 17:26:27", "content": "I mean before training I delete the boxes with very small or big area from the csv
", "rawMarkdown": "I mean before training I delete the boxes with very small or big area from the csv", "votes": null }, { "id": "893808", "postDate": "06/20/2020 00:07:44", "content": "What was your cutoff? There are very few boxes that can be excluded solely on area. I think I removed < 400 square pixels and > 145360. This is only a couple hundred boxes.
", "rawMarkdown": "What was your cutoff? There are very few boxes that can be excluded solely on area. I think I removed < 400 square pixels and > 145360. This is only a couple hundred boxes.", "votes": null }, { "id": "893866", "postDate": "06/20/2020 03:32:59", "content": "well my range was<200000 and > 100
", "rawMarkdown": "well my range was<200000 and > 100", "votes": null }, { "id": "894180", "postDate": "06/20/2020 07:59:43", "content": "I wonder which faster rcnn implementation you use, that's really cool
", "rawMarkdown": "I wonder which faster rcnn implementation you use, that's really cool", "votes": null }, { "id": "895599", "postDate": "06/21/2020 13:40:20", "content": "Include @mrinath as well. Do you guys only remove these pretty small & large bboxes only from the training set? Do we also need to remove them from validation set?
\n\nI'm new to this topic, so hope you can help me clarify. Thanks!
", "rawMarkdown": "Include @mrinath as well. Do you guys only remove these pretty small & large bboxes only from the training set? Do we also need to remove them from validation set?\n\nI'm new to this topic, so hope you can help me clarify. Thanks!", "votes": null }, { "id": "895745", "postDate": "06/21/2020 15:20:41", "content": "Well, I first removed all the adnormal sized boxes from csv then I split in train/validation.
", "rawMarkdown": "Well, I first removed all the adnormal sized boxes from csv then I split in train/validation.", "votes": null }, { "id": "903639", "postDate": "06/27/2020 02:37:06", "content": "Hi @mrinath, thanks for your reply and sorry for my late reply(VPN service is unstable recently). I tried the same strategy and my LB score drops though the local mAP increases...
", "rawMarkdown": "Hi @mrinath, thanks for your reply and sorry for my late reply(VPN service is unstable recently). I tried the same strategy and my LB score drops though the local mAP increases...", "votes": null } ], "comments": [ { "id": 892100, "author_name": "adizhol", "author_url": "", "post_date": "06/18/2020 16:47:56", "content": "What image resolution are you training?
", "votes": null, "replies": [ { "id": 892150, "author_name": "mrinath", "author_url": "", "post_date": "06/18/2020 17:19:48", "content": "on full image
", "votes": null, "replies": [] } ] }, { "id": 892227, "author_name": "matthewmasters", "author_url": "", "post_date": "06/18/2020 18:12:16", "content": "I’m at over 0.7 using fasterrcnn with resnet50. Single model, no TTA, no pseudolabeling.
\n\nJust have to keep experimenting
", "votes": null, "replies": [ { "id": 892247, "author_name": "mrinath", "author_url": "", "post_date": "06/18/2020 18:31:27", "content": "Have you used any ensembling techniques like wbf?
", "votes": null, "replies": [] }, { "id": 892320, "author_name": "matthewmasters", "author_url": "", "post_date": "06/18/2020 19:57:17", "content": "Yes, I've been using this since before it was made \"public\". Though that's not included in the >0.7 score I mentioned.
", "votes": null, "replies": [] }, { "id": 893420, "author_name": "mrinath", "author_url": "", "post_date": "06/19/2020 15:33:42", "content": "when I removed the boxes ( not that image) with abnormal sizes my lb score reduced., not getting why😩
", "votes": null, "replies": [] }, { "id": 893485, "author_name": "matthewmasters", "author_url": "", "post_date": "06/19/2020 16:40:04", "content": "What do you mean you remove boxes with abnormal sizes? From the training data? From your predictions?
", "votes": null, "replies": [] }, { "id": 893524, "author_name": "mrinath", "author_url": "", "post_date": "06/19/2020 17:26:27", "content": "I mean before training I delete the boxes with very small or big area from the csv
", "votes": null, "replies": [] }, { "id": 893808, "author_name": "matthewmasters", "author_url": "", "post_date": "06/20/2020 00:07:44", "content": "What was your cutoff? There are very few boxes that can be excluded solely on area. I think I removed < 400 square pixels and > 145360. This is only a couple hundred boxes.
", "votes": null, "replies": [] }, { "id": 893866, "author_name": "mrinath", "author_url": "", "post_date": "06/20/2020 03:32:59", "content": "well my range was<200000 and > 100
", "votes": null, "replies": [] }, { "id": 895599, "author_name": "", "author_url": "", "post_date": "06/21/2020 13:40:20", "content": "Include @mrinath as well. Do you guys only remove these pretty small & large bboxes only from the training set? Do we also need to remove them from validation set?
\n\nI'm new to this topic, so hope you can help me clarify. Thanks!
", "votes": null, "replies": [] }, { "id": 895745, "author_name": "mrinath", "author_url": "", "post_date": "06/21/2020 15:20:41", "content": "Well, I first removed all the adnormal sized boxes from csv then I split in train/validation.
", "votes": null, "replies": [] }, { "id": 903639, "author_name": "", "author_url": "", "post_date": "06/27/2020 02:37:06", "content": "Hi @mrinath, thanks for your reply and sorry for my late reply(VPN service is unstable recently). I tried the same strategy and my LB score drops though the local mAP increases...
", "votes": null, "replies": [] } ] }, { "id": 892556, "author_name": "nvnnghia", "author_url": "", "post_date": "06/19/2020 02:15:11", "content": "mine: 0.7314 no TTA, pseudo, wbf
", "votes": null, "replies": [ { "id": 892681, "author_name": "matthewmasters", "author_url": "", "post_date": "06/19/2020 04:38:51", "content": "Impressive! I take it you're back in? 😉
", "votes": null, "replies": [] }, { "id": 892771, "author_name": "nvnnghia", "author_url": "", "post_date": "06/19/2020 06:28:12", "content": "Hopefully I won't give up this time. :))
", "votes": null, "replies": [] }, { "id": 892870, "author_name": "mrinath", "author_url": "", "post_date": "06/19/2020 08:14:54", "content": "that's really impressive\nI will keep trying
", "votes": null, "replies": [] }, { "id": 894180, "author_name": "tiandaye", "author_url": "", "post_date": "06/20/2020 07:59:43", "content": "I wonder which faster rcnn implementation you use, that's really cool
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "891976": "So I have approached the problem using fastrcnn resnet101 backbone (since it was given better results than resnet50) and have applied different augmentation techniques with the **WBF** ensembling.\nStill, my score cant cross the 0.7 mark.\nSo I was thinking how much can you score just by using fastrcnn.\nI don't know why my score has got stagnant in a single value.", "892100": "What image resolution are you training?", "892150": "on full image", "892227": "I’m at over 0.7 using fasterrcnn with resnet50. Single model, no TTA, no pseudolabeling.\n\nJust have to keep experimenting", "892247": "Have you used any ensembling techniques like wbf?", "892320": "Yes, I've been using this since before it was made \"public\". Though that's not included in the >0.7 score I mentioned.", "892556": "mine: 0.7314 no TTA, pseudo, wbf", "892681": "Impressive! I take it you're back in? 😉", "892771": "Hopefully I won't give up this time. :))", "892870": "that's really impressive\nI will keep trying", "893420": "when I removed the boxes ( not that image) with abnormal sizes my lb score reduced., not getting why😩", "893485": "What do you mean you remove boxes with abnormal sizes? From the training data? From your predictions?", "893524": "I mean before training I delete the boxes with very small or big area from the csv", "893808": "What was your cutoff? There are very few boxes that can be excluded solely on area. I think I removed < 400 square pixels and > 145360. This is only a couple hundred boxes.", "893866": "well my range was<200000 and > 100", "894180": "I wonder which faster rcnn implementation you use, that's really cool", "895599": "Include @mrinath as well. Do you guys only remove these pretty small & large bboxes only from the training set? Do we also need to remove them from validation set?\n\nI'm new to this topic, so hope you can help me clarify. Thanks!", "895745": "Well, I first removed all the adnormal sized boxes from csv then I split in train/validation.", "903639": "Hi @mrinath, thanks for your reply and sorry for my late reply(VPN service is unstable recently). I tried the same strategy and my LB score drops though the local mAP increases..." }, "source": "meta" }