nID parameter for Darknet model initialization

Zhongdao / Towards-Realtime-MOT

Joint Detection and Embedding for fast multi-object tracking

MIT License

2.37k stars 541 forks source link

nID parameter for Darknet model initialization #110

Closed flysofast closed 4 years ago

flysofast commented 4 years ago

Hi, I'm trying to benchmark your jde.1088x608.uncertainty pre-trained model on CVPR19 dataset. I noticed that by default, your JDETracker initialize the Darknet model with nID = 14455. I have 3 questions:

Where does this value come from? is it the total number of object identities in the whole joint dataset that you used for training the model? because seemingly there are not that many identities in any of the datasets you mentioned in DATASET_ZOO.
- Is the value important to the performance of the model, since I could still run the tracking with fairly good result (~60% MOTA) with the same default nID value of 14455?
- If I want to train this model on a different dataset, how should I set this value for the model initialization?

Zhongdao commented 4 years ago

Yes, the nID parameter is supposed to be the total number of object identities in the whole joint dataset. For convenience (I am lazy), I use a large number 14455 since it is an estimation of the upper bound of the exact ID numbers. It seems better to use the exact ID number, but I think it is not a big problem if the number is larger than the exact nID. This is because the nID parameter is only used to initialize the cross-entropy classifier, I think if the classifier has some redundancy classes the performance will not be affected a lot.

Finally, I suggest you use the exact number of identities.

Zhongdao commented 4 years ago

By the way, do you test on the CVPR-19 training set or test set? Would you mind sharing me with your results? I would appreciate it. If you don't want to post it here in this issue, this is my email: wcd17@mails.tsinghua.edu.cn

Many thanks!

mheriyanto commented 3 years ago

Yes, the nID parameter is supposed to be the total number of object identities in the whole joint dataset. For convenience (I am lazy), I use a large number 14455 since it is an estimation of the upper bound of the exact ID numbers. It seems better to use the exact ID number, but I think it is not a big problem if the number is larger than the exact nID. This is because the nID parameter is only used to initialize the cross-entropy classifier, I think if the classifier has some redundancy classes the performance will not be affected a lot.

Finally, I suggest you use the exact number of identities.

Thanks for your explanation @Zhongdao