Poor training results on custom data with ''Temporal ROI Align for Video Object Recognition'' ?

[danial880]

I am training my custom dataset on video object detection. My dataset was in VOC format. I have written a custom script to convert VOC to COCOVid format. I have visualized the bounding boxes in my train and val jsons, they are correct. Here is the custom config file I used for training. The training logs show good accuracy but zero bounding box loss. I further dug into it and traced back to mmcv epoch-based runner for getting ground truth labels. The ground truth was not the same as in the JSON file. My annotations are already resized to the training image size (1000,1000). So how ground truth problem can be solved, please suggest.