search

ylabbe / cosypose

Code for "CosyPose: Consistent multi-view multi-object 6D pose estimation", ECCV 2020.
MIT License
301 stars 89 forks source link

Bounding Box Values In Pickle Files #60

Open azad96 opened 3 years ago

azad96 commented 3 years ago

I've created a pickle file, which has a compatible format to be called with the load_pix2pose_results(...) function, for a dataset that I've generated. After generating rgb images, scene_gt.json and scene_camera.json files with BlenderProc, I've also generated mask, mask_visib images and scene_gt_info.json file which includes "bbox_org" and "bbox_visib" information.

For example, bbox_org and bbox_visib are exactly same when all objects are completely visible as follows

image1_screenshot_09 06 2021

However, they are different when occlusion occurs as it can be seen below

image6_screenshot_09 06 2021

My question is which bounding box values should be used to create a pickle file for training? bbox_org or bbox_visib?

hannes56a commented 3 years ago

Hi, I currently trying to convert these pickle files I produced by making synthetic data ("run_dataset_recording.py") to "BOP format". Is there any help you provide? I'm able to get the rgb and the mask image (one object). But i'm currently helpless with the camera parameters and so on... I want to test (functionally) to train and test the detector on an syntethic only dataset. I use the "run_detector_training.py" and can train on the pcl files only, but for evaluation is the bop format needed... Perhaps someone can help me? I post the question to this issue, because its also according to the pickle files.

Thanks a lot!

azad96 commented 3 years ago

@hannes56a Are you trying to create a pickle file to be used in evaluation? load_pix2pose_results() function, for example, loads a pickle file to do evaluation?

hannes56a commented 3 years ago

@azad96 Thanks for the hint. I didn't try yet. I will have a look to this.