Training and Testing MOTRv2 on BDD100K Dataset?

[back2zack]

Hello,

First, thank you for your outstanding work on the MOTRv2 model. I have been experimenting with this model and am particularly interested in training and testing it on diffrent automotive datasets. it is mentioned in the paper that the model was trained on the BDD100K dataset.

However, I noticed that the repository does not include specific scripts or guidelines for training and evaluating the model on the BDD100K dataset. I am reaching out to ask if anyone in the community has successfully trained and tested MOTRv2 on BDD100K and would be willing to share their experience or provide any code snippets or detailed instructions.

Specifically, I am looking for:

• Preprocessing steps required for BDD100K.
• Training configurations and parameters.
• Evaluation scripts compatible with BDD100K.
• Any tips or best practices for optimizing performance on this dataset.

Any assistance or pointers to relevant resources would be greatly appreciated. Thank you in advance for your help!

Best regards,

[zzzz737]

I used bdd100k tracking data set, extracted "car" category from val data set to make truth value, and used part of it as training set:

1. For data sets without detection targets, I did not delete the frame directly, but modified the code to make its detection truth value null;
2. All parameters use the default parameters provided by the author to train 50 epochs;
3. The final loss hovers around 1.5

When testing: During testing, it was found that there were highly overlapping targets in the generated txt file, and the trackid had changed！ Do you know why that is！