{ "id": 307668, "title": "Does tracking help to boost private LB? ", "url": "/competitions/tensorflow-great-barrier-reef/discussion/307668", "author_name": "", "post_date": "2022-02-15T06:29:38.568969900Z", "votes": 5, "comment_count": 8, "views": 0, "content": "

First of all Congrats to all the winners and Great Competition, I learned a lot.

\n

Me and My team, Selected the following models.

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ModelPublic LB
2 ensemble notebooks with tracking0.684 - 0.682
1 ensemble[No tracking]0.672
solo model [No Tracking]0.674
\n

In Public LB, Ensemble with tracking, boost score (0.672 ---> 0.684).
\nBut in private LB Ensemble [No tracking] given the best score.

\n

I just wanted to ask does tracking help anyone to boost private score?
\nDoes testing data was not in sequence? Can we get test data after the competition ends?

", "messages": [ { "id": "1690887", "postDate": "02/15/2022 06:29:38", "content": "

First of all Congrats to all the winners and Great Competition, I learned a lot.

\n

Me and My team, Selected the following models.

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ModelPublic LB
2 ensemble notebooks with tracking0.684 - 0.682
1 ensemble[No tracking]0.672
solo model [No Tracking]0.674
\n

In Public LB, Ensemble with tracking, boost score (0.672 ---> 0.684).
\nBut in private LB Ensemble [No tracking] given the best score.

\n

I just wanted to ask does tracking help anyone to boost private score?
\nDoes testing data was not in sequence? Can we get test data after the competition ends?

", "rawMarkdown": "First of all Congrats to all the winners and Great Competition, I learned a lot. \n\nMe and My team, Selected the following models.\n\n| Model | Public LB |\n| --- | --- |\n| 2 ensemble notebooks with tracking | 0.684 - 0.682|\n| 1 ensemble[No tracking] | 0.672 |\n| solo model [No Tracking] | 0.674 |\n\nIn Public LB, Ensemble with tracking, boost score (0.672 ---> 0.684).\nBut in private LB Ensemble [No tracking] given the best score.\n\nI just wanted to ask does tracking help anyone to boost private score?\nDoes testing data was not in sequence? Can we get test data after the competition ends?", "votes": null }, { "id": "1690894", "postDate": "02/15/2022 06:34:13", "content": "

According to what I read, they will not be making private data public. They may be gathering more data (someone's going diving in the great barrier reef. Sounds like a nice vacation) and releasing it however.

\n

For me, tracking improved my private score. Not much, but it's an improvement.

\n

I'm 99% certain that the private test was in sequence. There might be small bugs (there's one tiny one in the public tracking notebook)

", "rawMarkdown": "According to what I read, they will not be making private data public. They may be gathering more data (someone's going diving in the great barrier reef. Sounds like a nice vacation) and releasing it however.\n\nFor me, tracking improved my private score. Not much, but it's an improvement.\n\nI'm 99% certain that the private test was in sequence. There might be small bugs (there's one tiny one in the public tracking notebook)", "votes": null }, { "id": "1690904", "postDate": "02/15/2022 06:41:48", "content": "

Yes it does, you can try it on the 0.665 notebook from @freshair1996
\nThere's a 0.02 gain on both public and private score
\nIt was ensembling that does not give gain on my case

", "rawMarkdown": "Yes it does, you can try it on the 0.665 notebook from @freshair1996 \nThere's a 0.02 gain on both public and private score\nIt was ensembling that does not give gain on my case", "votes": null }, { "id": "1690947", "postDate": "02/15/2022 07:02:41", "content": "

I guess Tracking worked well on solo model or high conf submissions.

", "rawMarkdown": "I guess Tracking worked well on solo model or high conf submissions.", "votes": null }, { "id": "1692150", "postDate": "02/15/2022 21:30:07", "content": "

If the popular public notebook here doesn't help your private score, try adding the second line below. Then tracking will be more conservative and only add cots if they were in previous frame but missing in current frame.

\n
if last_detected_frame_id == frame_id: continue\nif last_detected_frame_id <= frame_id-2:  continue\n
", "rawMarkdown": "If the popular public notebook [here][1] doesn't help your private score, try adding the second line below. Then tracking will be more conservative and only add cots if they were in previous frame but missing in current frame.\n\n if last_detected_frame_id == frame_id: continue\n if last_detected_frame_id <= frame_id-2: continue\n\n[1]: https://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642", "votes": null }, { "id": "1693880", "postDate": "02/17/2022 03:18:35", "content": "

@cdeotte With that line, my private score jump to 685 and I would have gotten a silver 👀

\n

https://www.kaggle.com/nyleve/silver-public-model-goodmoon-tracking-aug-true

", "rawMarkdown": "cdeotte With that line, my private score jump to 685 and I would have gotten a silver 👀\n\nhttps://www.kaggle.com/nyleve/silver-public-model-goodmoon-tracking-aug-true", "votes": null }, { "id": "1693893", "postDate": "02/17/2022 03:25:51", "content": "

Awesome!

", "rawMarkdown": "Awesome!", "votes": null }, { "id": "1693947", "postDate": "02/17/2022 04:35:20", "content": "

@cdeotte,
\nThanks for the trick, we used global-tracking-method-homography method for tracking,
\nBut I got my answer, Our model was a little overfitted and the wrong CONF score for the tracker is the reason we missed the bronze area.

", "rawMarkdown": "cdeotte, \nThanks for the trick, we used [global-tracking-method-homography](https://www.kaggle.com/thomasdubail/global-tracking-method-homography) method for tracking, \nBut I got my answer, Our model was a little overfitted and the wrong CONF score for the tracker is the reason we missed the bronze area.", "votes": null }, { "id": "1695876", "postDate": "02/18/2022 12:23:33", "content": "

it added 0.007 to private and 0.002 to public, overall 23rd with it 8 places up! :)

", "rawMarkdown": "it added 0.007 to private and 0.002 to public, overall 23rd with it 8 places up! :)", "votes": null } ], "comments": [ { "id": 1690894, "author_name": "liminchen1", "author_url": "", "post_date": "02/15/2022 06:34:13", "content": "

According to what I read, they will not be making private data public. They may be gathering more data (someone's going diving in the great barrier reef. Sounds like a nice vacation) and releasing it however.

\n

For me, tracking improved my private score. Not much, but it's an improvement.

\n

I'm 99% certain that the private test was in sequence. There might be small bugs (there's one tiny one in the public tracking notebook)

", "votes": null, "replies": [] }, { "id": 1690904, "author_name": "nyleve", "author_url": "", "post_date": "02/15/2022 06:41:48", "content": "

Yes it does, you can try it on the 0.665 notebook from @freshair1996
\nThere's a 0.02 gain on both public and private score
\nIt was ensembling that does not give gain on my case

", "votes": null, "replies": [ { "id": 1690947, "author_name": "mahipalsingh", "author_url": "", "post_date": "02/15/2022 07:02:41", "content": "

I guess Tracking worked well on solo model or high conf submissions.

", "votes": null, "replies": [] } ] }, { "id": 1692150, "author_name": "cdeotte", "author_url": "", "post_date": "02/15/2022 21:30:07", "content": "

If the popular public notebook here doesn't help your private score, try adding the second line below. Then tracking will be more conservative and only add cots if they were in previous frame but missing in current frame.

\n
if last_detected_frame_id == frame_id: continue\nif last_detected_frame_id <= frame_id-2:  continue\n
", "votes": null, "replies": [ { "id": 1693880, "author_name": "nyleve", "author_url": "", "post_date": "02/17/2022 03:18:35", "content": "

@cdeotte With that line, my private score jump to 685 and I would have gotten a silver 👀

\n

https://www.kaggle.com/nyleve/silver-public-model-goodmoon-tracking-aug-true

", "votes": null, "replies": [] }, { "id": 1693893, "author_name": "cdeotte", "author_url": "", "post_date": "02/17/2022 03:25:51", "content": "

Awesome!

", "votes": null, "replies": [] }, { "id": 1693947, "author_name": "mahipalsingh", "author_url": "", "post_date": "02/17/2022 04:35:20", "content": "

@cdeotte,
\nThanks for the trick, we used global-tracking-method-homography method for tracking,
\nBut I got my answer, Our model was a little overfitted and the wrong CONF score for the tracker is the reason we missed the bronze area.

", "votes": null, "replies": [] }, { "id": 1695876, "author_name": "lukaszborecki", "author_url": "", "post_date": "02/18/2022 12:23:33", "content": "

it added 0.007 to private and 0.002 to public, overall 23rd with it 8 places up! :)

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1690887": "First of all Congrats to all the winners and Great Competition, I learned a lot. \n\nMe and My team, Selected the following models.\n\n| Model | Public LB |\n| --- | --- |\n| 2 ensemble notebooks with tracking | 0.684 - 0.682|\n| 1 ensemble[No tracking] | 0.672 |\n| solo model [No Tracking] | 0.674 |\n\nIn Public LB, Ensemble with tracking, boost score (0.672 ---> 0.684).\nBut in private LB Ensemble [No tracking] given the best score.\n\nI just wanted to ask does tracking help anyone to boost private score?\nDoes testing data was not in sequence? Can we get test data after the competition ends?", "1690894": "According to what I read, they will not be making private data public. They may be gathering more data (someone's going diving in the great barrier reef. Sounds like a nice vacation) and releasing it however.\n\nFor me, tracking improved my private score. Not much, but it's an improvement.\n\nI'm 99% certain that the private test was in sequence. There might be small bugs (there's one tiny one in the public tracking notebook)", "1690904": "Yes it does, you can try it on the 0.665 notebook from @freshair1996 \nThere's a 0.02 gain on both public and private score\nIt was ensembling that does not give gain on my case", "1690947": "I guess Tracking worked well on solo model or high conf submissions.", "1692150": "If the popular public notebook [here][1] doesn't help your private score, try adding the second line below. Then tracking will be more conservative and only add cots if they were in previous frame but missing in current frame.\n\n if last_detected_frame_id == frame_id: continue\n if last_detected_frame_id <= frame_id-2: continue\n\n[1]: https://www.kaggle.com/kocha1/only-yolov5-tracking-lb-642", "1693880": "cdeotte With that line, my private score jump to 685 and I would have gotten a silver 👀\n\nhttps://www.kaggle.com/nyleve/silver-public-model-goodmoon-tracking-aug-true", "1693893": "Awesome!", "1693947": "cdeotte, \nThanks for the trick, we used [global-tracking-method-homography](https://www.kaggle.com/thomasdubail/global-tracking-method-homography) method for tracking, \nBut I got my answer, Our model was a little overfitted and the wrong CONF score for the tracker is the reason we missed the bronze area.", "1695876": "it added 0.007 to private and 0.002 to public, overall 23rd with it 8 places up! :)" }, "source": "meta" }