Hi CAMELTrack authors,
Thank you for the excellent work and for releasing the code. It has been very helpful for our research.
We are currently working on SportsMOT and trying to set up fair comparisons, but we are a bit uncertain about the training splits used for the reported SportsMOT test results.
After checking the paper, supplementary material, and released code, we would like to clarify whether the CAMEL association module was trained using only the training set, or using the combined train+val set.
We noticed that the supplementary material mentions that the ReID embeddings for the test set are generated with a model jointly trained on train+val. We also noticed a comment in the SportsMOT dataset wrapper in TrackLab indicating that some detection files were generated with YOLOX weights trained on train+val. We understand these may refer only to the detector/ReID components, but it would be very helpful to know the exact training splits used for each component.
This information would help us reproduce and compare results under a consistent setting. We will make sure to cite CAMELTrack in our related research.
Thank you very much for your clarification.
Hi CAMELTrack authors,
Thank you for the excellent work and for releasing the code. It has been very helpful for our research.
We are currently working on SportsMOT and trying to set up fair comparisons, but we are a bit uncertain about the training splits used for the reported SportsMOT test results.
After checking the paper, supplementary material, and released code, we would like to clarify whether the CAMEL association module was trained using only the training set, or using the combined train+val set.
We noticed that the supplementary material mentions that the ReID embeddings for the test set are generated with a model jointly trained on train+val. We also noticed a comment in the SportsMOT dataset wrapper in TrackLab indicating that some detection files were generated with YOLOX weights trained on train+val. We understand these may refer only to the detector/ReID components, but it would be very helpful to know the exact training splits used for each component.
This information would help us reproduce and compare results under a consistent setting. We will make sure to cite CAMELTrack in our related research.
Thank you very much for your clarification.