It is clearly known factor that there many noisy bbx in the train data unlike test data.
\nHere are things that worked best for me in 10 days
1) Label correction of small bounding boxes by decreasing the x,y with 2 percent and increasing width ,height by factor of 1.5 . This would have enlarged very point size bbx to some visible level without affecting much accuracy of bboxes
\n2) Removing all bbx with area less than 500 after above correction
\n3) Pseudo label training with img size 1024 by removing the noisy train images and usign it along with test images which were far more accurate.
\n4) 512 img size model worked best on Public LB but it was 1024 that worked best on private LB,it is true in atleast my case.
\nThese above were key things i could think of at best in last 10 days.
\nI regret i dint select the submission of 1024 img size worked based on above observation and it apparently turned to be one scoring highest in Private lb but moderately on public lb. 67.13 unselected
Few other things due short of time i couldnt explore was Faster RCNN with mobile Net v2 , this was light model could have worked in this competition.
\nCongrats to all top positions..
\nMay be if i had joined earlier could have been better placed.
\nNext Competition is OSIC…
", "messages": [ { "id": 958611, "postDate": "2020-08-05T04:41:04.117Z", "content": "It is clearly known factor that there many noisy bbx in the train data unlike test data.
\nHere are things that worked best for me in 10 days
1) Label correction of small bounding boxes by decreasing the x,y with 2 percent and increasing width ,height by factor of 1.5 . This would have enlarged very point size bbx to some visible level without affecting much accuracy of bboxes
\n2) Removing all bbx with area less than 500 after above correction
\n3) Pseudo label training with img size 1024 by removing the noisy train images and usign it along with test images which were far more accurate.
\n4) 512 img size model worked best on Public LB but it was 1024 that worked best on private LB,it is true in atleast my case.
\nThese above were key things i could think of at best in last 10 days.
\nI regret i dint select the submission of 1024 img size worked based on above observation and it apparently turned to be one scoring highest in Private lb but moderately on public lb. 67.13 unselected
Few other things due short of time i couldnt explore was Faster RCNN with mobile Net v2 , this was light model could have worked in this competition.
\nCongrats to all top positions..
\nMay be if i had joined earlier could have been better placed.
\nNext Competition is OSIC…
", "rawMarkdown": "It is clearly known factor that there many noisy bbx in the train data unlike test data. \nHere are things that worked best for me in 10 days\n\n1) Label correction of small bounding boxes by decreasing the x,y with 2 percent and increasing width ,height by factor of 1.5 . This would have enlarged very point size bbx to some visible level without affecting much accuracy of bboxes\n\n2) Removing all bbx with area less than 500 after above correction\n\n3) Pseudo label training with img size 1024 by removing the noisy train images and usign it along with test images which were far more accurate. \n\n4) 512 img size model worked best on Public LB but it was 1024 that worked best on private LB,it is true in atleast my case.\n\nThese above were key things i could think of at best in last 10 days. \nI regret i dint select the submission of 1024 img size worked based on above observation and it apparently turned to be one scoring highest in Private lb but moderately on public lb. 67.13 unselected \n\nFew other things due short of time i couldnt explore was Faster RCNN with mobile Net v2 , this was light model could have worked in this competition. \n\nCongrats to all top positions..\n\nMay be if i had joined earlier could have been better placed.\n\nNext Competition is OSIC..." }, { "id": 1338988, "postDate": "2021-06-06T23:38:14.357Z", "content": "It helps to understand how kaggle works
", "rawMarkdown": "It helps to understand how kaggle works " }, { "id": 1338777, "postDate": "2021-06-06T18:12:32.467Z", "content": "Here is unslected gold kernel https://www.kaggle.com/jaideepvalani/normal-pl-eff-tta-wbf-1024/data?scriptVersionId=40158023
", "rawMarkdown": "Here is unslected gold kernel https://www.kaggle.com/jaideepvalani/normal-pl-eff-tta-wbf-1024/data?scriptVersionId=40158023" }, { "id": 958731, "postDate": "2020-08-05T06:00:39.487Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 958996, "postDate": "2020-08-05T09:16:59.837Z", "content": "thanks yeah.. sole objective was getting acquainted with obj detection which i never worked before ,just had super facial knowledge to it .
", "rawMarkdown": "thanks yeah.. sole objective was getting acquainted with obj detection which i never worked before ,just had super facial knowledge to it . \n " } ] } ], "comments": [ { "id": 1338988, "author_name": "assign", "author_url": "", "post_date": "2021-06-06T23:38:14.357000", "content": "It helps to understand how kaggle works
", "votes": 0, "replies": [] }, { "id": 1338777, "author_name": "Jaideep", "author_url": "", "post_date": "2021-06-06T18:12:32.467000", "content": "Here is unslected gold kernel https://www.kaggle.com/jaideepvalani/normal-pl-eff-tta-wbf-1024/data?scriptVersionId=40158023
", "votes": 0, "replies": [] }, { "id": 958731, "author_name": "", "author_url": "", "post_date": "2020-08-05T06:00:39.487000", "content": "", "votes": 0, "replies": [ { "id": 958996, "author_name": "Jaideep", "author_url": "", "post_date": "2020-08-05T09:16:59.837000", "content": "thanks yeah.. sole objective was getting acquainted with obj detection which i never worked before ,just had super facial knowledge to it .
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "958611": "It is clearly known factor that there many noisy bbx in the train data unlike test data. \nHere are things that worked best for me in 10 days\n\n1) Label correction of small bounding boxes by decreasing the x,y with 2 percent and increasing width ,height by factor of 1.5 . This would have enlarged very point size bbx to some visible level without affecting much accuracy of bboxes\n\n2) Removing all bbx with area less than 500 after above correction\n\n3) Pseudo label training with img size 1024 by removing the noisy train images and usign it along with test images which were far more accurate. \n\n4) 512 img size model worked best on Public LB but it was 1024 that worked best on private LB,it is true in atleast my case.\n\nThese above were key things i could think of at best in last 10 days. \nI regret i dint select the submission of 1024 img size worked based on above observation and it apparently turned to be one scoring highest in Private lb but moderately on public lb. 67.13 unselected \n\nFew other things due short of time i couldnt explore was Faster RCNN with mobile Net v2 , this was light model could have worked in this competition. \n\nCongrats to all top positions..\n\nMay be if i had joined earlier could have been better placed.\n\nNext Competition is OSIC...", "1338988": "It helps to understand how kaggle works ", "1338777": "Here is unslected gold kernel https://www.kaggle.com/jaideepvalani/normal-pl-eff-tta-wbf-1024/data?scriptVersionId=40158023", "958731": "" } }