Evaluating on custom dataset

ylabbe / cosypose

Code for "CosyPose: Consistent multi-view multi-object 6D pose estimation", ECCV 2020.

MIT License

301 stars 89 forks source link

Evaluating on custom dataset #65

Closed hannes56a closed 2 years ago

hannes56a commented 2 years ago

Hi all, I would like to evaluate on a custom dataset. I tried to use the "run_cosypose_eval.py", but struggled. Are there any hints about that?

hannes56a commented 2 years ago

Hey, its me again. While evaluation (run_cosypose_eval.py" I need a *.pkl file (for example tless located in "saved_detections"). For eval on my custom dataset i do not have such a pkl file. Can someone help me please?

Why does I need this file? I thougt the groundtruth is in the "scene_gt.json" !?
How can I generate such a file? Is there some script? Does anyone did it allready?
What is in this file? I think its the groundtruth?

FriendlyNature commented 2 years ago

Hi, my assumption is, that the pkl file contains the 2D bounding box detections of an independent detection model. Therefore, it is not ground truth data but contains predictions that might be faulty. However, I would also be interested in generating such an file for a custom dataset. Has someone an answer to @hannes56a second question:

How can I generate such a file? Is there some script? Does anyone did it allready?