Detections with own dataset

[caarmeecoorbii]

Hello, thank you for your repository as it has been very helpful for my final degree project. I tried your method with my own dataset and it really improves the results a lot. Now, I want to further improve my results. Using the 'generate_detections.py' file, I am able to generate detection features based on FastReID. This gives me a .npy file for each sequence. I would like to edit this .npy file to add more detections separately. Could you tell me the structure followed by these .npy files so that I can add additional detections?"

Thanks!

Carme

[dyhBUPT]

Hi, glad to hear that StrongSORT works in your case.
For your question, please refer to this issue #28.

[caarmeecoorbii]

Hi, thank you for providing this information. I have a question regarding the 2048 features mentioned in issue #28. Are these features the result of applying FastReID? I'm asking because I've generated detections using other methods to potentially improve results in StrongSORT.

Now, I want to add these detections to the .npy file, but I'm unsure about obtaining features for these detections. How would I go about passing them to FastReID for feature extraction?

In the generate_detections.py file, I know it generates these features from the input detection .txt files. So, would it be possible for me to edit this script to accommodate the new detections?

Carme Corbi

[dyhBUPT]

Yes, these features are from FastReID. So please modify the generate_detections.py for your requirements.

[caarmeecoorbii]

I have another question. If I want to improve my results obtained previously with the SORT method, is there a way to train the StrongSORT method, or do you recommend that I modify some parts of the code to achieve better tracking? Additionally, I've noticed that some images do not detect all the people, and I would like to improve this aspect as well.

Thank you!

[dyhBUPT]

If you have "off-the-shelf tracking results" (e.g., multiple txt files), and want to improve the performance upon it, it is recommended to try some "offline post-processing" methods, e.g., global link, split-and-connect, interpolation, etc.

BTW, why not directly inputting the raw detections and features into StrongSORT? SORT is an outdated method.

[caarmeecoorbii]

Thank you for your reply. Do you know any post processing methods that get good results for MOT? Sorry for the confusion, I meant StrongSORT.

[dyhBUPT]

For post-processing, this repo supports interpolation and appearance-free global link.
Indeed, you can try appearance-based link, which is expected to achieve better results.
Based on StrongSORT, you can also try to borrow other plug-and-play tricks from other SOTA works, e.g., BYTE.

[caarmeecoorbii]

Thank you so much! Yes, I tried AFLink and it improves my results. But still object association could be improved as many new ids are assigned for players. Also, in the dataset that I am using, which are of soccer players, the detections when they are in a lot of movement are not accurate.

[dyhBUPT]

Maybe you can try to release the motion constraints.

[caarmeecoorbii]

Apologies, but which part of the code are you referring to?

[dyhBUPT]

All parts related to motion cost and matching can be considered.

[caarmeecoorbii]

do you think if I modify the hyperparameters I could get better results?

[dyhBUPT]

yes, it usually works.

…

--------------原始邮件-------------- 发件人："caarmeecoorbii ***@***.***>; 发送时间：2024年3月30日(星期六) 下午3:06 收件人："dyhBUPT/StrongSORT" ***@***.***>; 抄送："dyhBUPT ***@***.***>;"Comment ***@***.***>; 主题：Re: [dyhBUPT/StrongSORT] Detections with own dataset (Issue #107) ----------------------------------- do you think if I modify the hyperparameters I could get better results? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[caarmeecoorbii]

I tried changing the value of EMA_alpha to 0.8 with this command :
python strong_sort.py MOT16 test --BoT --EMA --EMA_alpha 0.8 --NSA --MC --woC --AFLink but I get the following error:

[dyhBUPT]

Oh, to use --EMA_alpha in the command line, you should specify the type of --EMA_alpha to float in opt.py.

[caarmeecoorbii]

Apologies for the digression, but I was wondering if you happen to know whether DeepSORT tracker can be trained? I've modified the parameters of StrongSORT and managed to achieve slightly better results, but the ID assignment still isn't quite accurate. Therefore, I'd like to train the DeepSORT tracker if possible, as StrongSORT cannot be trained.

[dyhBUPT]

In DeepSORT, only the detector and ReID model can be trained.

…

--------------原始邮件-------------- 发件人："caarmeecoorbii ***@***.***>; 发送时间：2024年4月5日(星期五) 凌晨1:45 收件人："dyhBUPT/StrongSORT" ***@***.***>; 抄送："dyhBUPT ***@***.***>;"Comment ***@***.***>; 主题：Re: [dyhBUPT/StrongSORT] Detections with own dataset (Issue #107) ----------------------------------- Apologies for the digression, but I was wondering if you happen to know whether DeepSORT tracker can be trained? I've modified the parameters of StrongSORT and managed to achieve slightly better results, but the ID assignment still isn't quite accurate. Therefore, I'd like to train the DeepSORT tracker if possible, as StrongSORT cannot be trained. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[caarmeecoorbii]

How can I train the reID model for DeepSORT with my own dataset?

[dyhBUPT]

FastReID is recommended.

…

--------------原始邮件-------------- 发件人："caarmeecoorbii ***@***.***>; 发送时间：2024年4月6日(星期六) 下午5:06 收件人："dyhBUPT/StrongSORT" ***@***.***>; 抄送："dyhBUPT ***@***.***>;"Comment ***@***.***>; 主题：Re: [dyhBUPT/StrongSORT] Detections with own dataset (Issue #107) ----------------------------------- How can I train the reID model for DeepSORT with my own dataset? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>