Can you provide the specific data formats of the results?(the orientation of bev mask, the facing direction of car and something more ...). It would help a lot if you can provide a new result.zip file(on test dataset) as a template. Thanks a lot.
We also found an extreme fall back on test set in terms of segmentation task, while the detection task share the similar performance with the val set. Wierd lol