Unexpected behavior when evaluation based on bounding box predictions

[ascCognify]

📚 The doc issue

When evaluating Top-Down models using bounding box predictions generated by an Object Detector, the following happens during validation/test phase:

1. No keypoints are visualized in the ground truth images.
2. The AP Coco Metric is close to 0.

To reproduce this problem, you can simply use the td-hm_ViTPose-large_8xb64-210e_coco-256x192.py config file. When setting bbox_file=None, everything works as expected (visible keypoints and reasonable AP scores).

What is the expected behavior of setting a bbox_file?

Suggest a potential alternative/fix

No response