![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1158131%2Fff468245fd620201bdeb0bf2e2879b0d%2F51b3e36ab.jpg?generation=1595661036941151&alt=media\")
Hi All,
\n\nI would like to know your thoughts about WBF and NMS wrt this competition.
\n\nFor comparison, I applied NMS and then WBF on test images one by one and when I compared the images with each approach I found that the confidence score with NMS is far better than WBF but after submission, I got score 0.6626 with NMS and 0.6815 with WBF.
\n\nImage with NMS-\n
Image with WBF-\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1158131%2F9be382015bcad176ff8d6557df60d11a%2F51b3e36ab.jpg?generation=1595661071702535&alt=media\")
Although WBF performance is better in this competition yet it is tested as a poor performer with Open Images dataset. Check this poor performance of WBF
\n\nI doubt the score is better with WBF is due to the scoring mechanism of this competition. What do you think? Will you use WBF outside of this- in your projects?
", "messages": [ { "id": "944546", "postDate": "07/25/2020 07:21:27", "content": "Hi All,
\n\nI would like to know your thoughts about WBF and NMS wrt this competition.
\n\nFor comparison, I applied NMS and then WBF on test images one by one and when I compared the images with each approach I found that the confidence score with NMS is far better than WBF but after submission, I got score 0.6626 with NMS and 0.6815 with WBF.
\n\nImage with NMS-\n
Image with WBF-\n
Although WBF performance is better in this competition yet it is tested as a poor performer with Open Images dataset. Check this poor performance of WBF
\n\nI doubt the score is better with WBF is due to the scoring mechanism of this competition. What do you think? Will you use WBF outside of this- in your projects?
", "rawMarkdown": "Hi All,\n\nI would like to know your thoughts about WBF and NMS wrt this competition.\n\nFor comparison, I applied NMS and then WBF on test images one by one and when I compared the images with each approach I found that the confidence score with NMS is far better than WBF but after submission, I got score 0.6626 with NMS and 0.6815 with WBF.\n\nImage with NMS-\n\n\nImage with WBF-\n\n\nAlthough WBF performance is better in this competition yet it is tested as a poor performer with Open Images dataset. Check this [poor performance of WBF](https://github.com/ZFTurbo/Weighted-Boxes-Fusion/issues/2)\n\nI doubt the score is better with WBF is due to the scoring mechanism of this competition. What do you think? Will you use WBF outside of this- in your projects?", "votes": null }, { "id": "945923", "postDate": "07/26/2020 08:11:01", "content": "Even if the confidence score with NMS is far \"better\" than WBF that does not mean that NMS should lead to better score for the simple reason that prediction confidence has almost no impact on the evaluation score. You can check how the evaluation score is measured for more details.
", "rawMarkdown": "Even if the confidence score with NMS is far \"better\" than WBF that does not mean that NMS should lead to better score for the simple reason that prediction confidence has almost no impact on the evaluation score. You can check how the evaluation score is measured for more details.", "votes": null }, { "id": "946027", "postDate": "07/26/2020 09:48:16", "content": "Yes, that I also got to know after my submission and comparison.
", "rawMarkdown": "Yes, that I also got to know after my submission and comparison.", "votes": null } ], "comments": [ { "id": 945923, "author_name": "chihaoui", "author_url": "", "post_date": "07/26/2020 08:11:01", "content": "Even if the confidence score with NMS is far \"better\" than WBF that does not mean that NMS should lead to better score for the simple reason that prediction confidence has almost no impact on the evaluation score. You can check how the evaluation score is measured for more details.
", "votes": null, "replies": [ { "id": 946027, "author_name": "iamprateek", "author_url": "", "post_date": "07/26/2020 09:48:16", "content": "Yes, that I also got to know after my submission and comparison.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "944546": "Hi All,\n\nI would like to know your thoughts about WBF and NMS wrt this competition.\n\nFor comparison, I applied NMS and then WBF on test images one by one and when I compared the images with each approach I found that the confidence score with NMS is far better than WBF but after submission, I got score 0.6626 with NMS and 0.6815 with WBF.\n\nImage with NMS-\n\n\nImage with WBF-\n\n\nAlthough WBF performance is better in this competition yet it is tested as a poor performer with Open Images dataset. Check this [poor performance of WBF](https://github.com/ZFTurbo/Weighted-Boxes-Fusion/issues/2)\n\nI doubt the score is better with WBF is due to the scoring mechanism of this competition. What do you think? Will you use WBF outside of this- in your projects?", "945923": "Even if the confidence score with NMS is far \"better\" than WBF that does not mean that NMS should lead to better score for the simple reason that prediction confidence has almost no impact on the evaluation score. You can check how the evaluation score is measured for more details.", "946027": "Yes, that I also got to know after my submission and comparison." }, "source": "meta" }