You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

YAML Metadata Warning: empty or missing yaml metadata in repo card (https://huggingface.co/docs/hub/datasets-cards)

Content

Each subdirectory represents a scene. Per scene we have:

<scene_id>_annotations.json: Segmentations
<scene_id>_descriptions.json: Natural-language activity descriptions
<scene_id>_laser_scan.ply: The scan mesh (point cloud + triangles)
<scene_id>_motions.json: The articulation annotations
<scene_id>_transform.npy: The transform matrix for going from laser_scan.ply to gaussians ply
point_cloud_30000.ply: The gaussians, 1.5M

For consolidated annotations: across all scenes, all annotation types (descriptions, articulations and obb): scenefun_single_tasks_obb.json

Important: The annotation .jsons are all on the gaussian laser scan. They need to be transformed via the .npy to apply to the gaussians. The reason for the misalignment is that the gaussians are constructed through a video, thus the video leads to different results (different rotaion and translation of the scene). That's why in the consolidated annotations we have both fields "obb [pos_x,_y,_z,angle_z,scale_x,_y,_z]" and "obb_laser_scan". The first one is already projected to gaussians, the second one is on the scan mesh.

Downloads last month: 3