Dear Authors,

I hope this email finds you well.

Please correct me if I am wrong, but based on my understanding of the code, all the "evaluate" function used in "main.py" require Hungarian Matching (from the SetCriterion function). This approach requires prior knowledge of the object classes in the /coco2017/val2017 dataset to select the appropriate queries from the 300 available queries. However, this assumption may not be compatible with real-world inference scenarios, where no prior information about the test data is available.

This approach ensures that CL-DETR always classifies detected objects with its correct classes.

Ideally, CLDETR should detect object boxes and their corresponding classes independently and then evaluate the results against the ground truth.

Given this, could you please revise the evaluation function to ensure it aligns with practical inference conditions?