[Feature] Using CustomDataset in Nuscenes style - How to support NuScenes style evaluation ?

[AlexIlis]

Custom Dataset Evaluation with NuScenes when certain dataset related jsons are non-existent

One of the ways to adapt our custom dataset into NuScenes style is by generating the middle format nuscnes_infos_train/val.trainval.pkl as expected for NuScenes. I have been succesful in doing that.

However, Evaluation on this custom Dataset is tricky as it compares predictions with GT from v1.0-trainval/sample_annotations.json which stores GT as translation:[], size:[] and rotation:[] unlike the 7DOF gt_bboxes that is available in infos.pkl

What is the best way to adapt evaluation on Nuscenes style infos.pkl with predictions?

Highlight:

context

Our custom dataset will not contain NuScenes like schema.