facebookresearch / co3d

Tooling for the Common Objects In 3D dataset.
Other
921 stars 71 forks source link

Filter accurate pointclouds #75

Open theamaya opened 1 year ago

theamaya commented 1 year ago

Hello,

Your paper mentions that a set of accurate pointclouds for each category has been selected. Where can I find the labels for these accurate pointclouds that passed your filtering?

davnov134 commented 1 year ago

Hi, each point cloud has a point cloud quality score. You can find these scores in the sequence_annotation.jgz files. To get the best pointclouds, I would suggest taking the ones with the highest scores.

theamaya commented 1 year ago

Thank you for your reply. However, even after downloading the entire dataset, the sequence_annotation.jgz files for each category appear to contain details for very few sequences per category.

Eg:

for 'backpack' category - {"sequence_name": "537_78249_152122", "category": "backpack", "video": null, "point_cloud": null, "viewpoint_quality_score": NaN}

would you please tell me what am I missing?

davnov134 commented 1 year ago

This sounds weird, are you sure that you downloaded the entire co3d? If yes, can you run the following example and check that the loaded sequence annotations contain many entries for backpacks? https://github.com/facebookresearch/co3d/blob/main/examples/print_co3d_stats.py

You can also run the following example to show backpack sequences: https://github.com/facebookresearch/co3d/blob/main/examples/show_co3d_dataset.py

jh27kim commented 8 months ago

Hi, how about the camera pose scores ? Is the higher the better ? Also, how well do the scores of point cloud correlate with camera pose scores ?

shapovalov commented 8 months ago

The scores are indeed the higher the better. They were both regressed based on COLMAP outputs, so probably do correlate.

jh27kim commented 8 months ago

Right, thank you for clarification.