graspnet_example/README.md

GraspNet-1Billion example scene (scene_0000, RealSense, all 256 views)

One complete scene of GraspNet-1Billion (Fang et al., CVPR 2020) with everything needed to reconstruct it: the 9 object meshes, every per-view object pose, the camera trajectory, and the RGB-D images with segmentation labels. It is included as the reference for rendering multi-object scenes in the GraspNet style.

Licence: this folder is CC BY-NC-SA 4.0 (GraspNet team, MVIG, SJTU: free for non-commercial use, redistribution under the same conditions, https://graspnet.net/about), separate from the CC-BY / CC0 objects in objects/. Cite:

@inproceedings{fang2020graspnet, title={GraspNet-1Billion: A Large-Scale Benchmark for General Object Grasping},
  author={Fang, Hao-Shu and Wang, Chenxi and Gou, Minghao and Lu, Cewu}, booktitle={CVPR}, year={2020}}

Files

scene_0000_realsense.zip
  scene_0000/object_id_list.txt            the 9 GraspNet object ids in this scene (14 0 5 15 20 48 66 70 46)
  scene_0000/rs_wrt_kn.npy                 RealSense pose with respect to the Kinect (not needed here)
  scene_0000/realsense/camK.npy            3x3 intrinsics (also inside every meta file)
  scene_0000/realsense/camera_poses.npy    (256,4,4) pose of camera v with respect to camera 0 (view 0)
  scene_0000/realsense/cam0_wrt_table.npy  (4,4) pose of camera 0 with respect to the table frame (z up)
  scene_0000/realsense/rgb/VVVV.png        1280x720 colour
  scene_0000/realsense/depth/VVVV.png      16-bit depth in millimetres (divide by factor_depth = 1000)
  scene_0000/realsense/label/VVVV.png      per-pixel GraspNet object id + 1, 0 = background
  scene_0000/realsense/meta/VVVV.mat       poses (3,4,N) object->camera transforms in metres, cls_indexes (N,) = object id + 1,
                                           intrinsic_matrix (3,3), factor_depth
  scene_0000/realsense/annotations/VVVV.xml the same poses as position + quaternion in the camera frame (graspnetAPI's XML form)
models_scene_0000.zip
  models/OID/textured.obj + .mtl + texture image, nontextured.ply, nontextured_simplified.ply   (metres, object frame)
reconstruct_scene.py                       standalone reconstruction and check (numpy, scipy, Pillow, trimesh)

Views are indexed 0..255 along the recorded camera trajectory; object poses are given per view in that view's camera frame, so the same scene can be reconstructed from any view. camera_poses.npy and cam0_wrt_table.npy put all views in one table frame: cam_v_wrt_table = cam0_wrt_table @ camera_poses[v].

Reconstruct

unzip scene_0000_realsense.zip; unzip models_scene_0000.zip
python reconstruct_scene.py --scene scene_0000 --models models --view 0 --out out            # camera frame of view 0
python reconstruct_scene.py --scene scene_0000 --models models --view 137 --table-frame --out out

The script prints each object's pose, writes out/scene_VVVV.glb (the textured meshes at their poses; add --no-texture for a small file) and out/depth_VVVV.ply (the back-projected depth of the object pixels, coloured), and reports the distance between those depth points and the reconstructed surfaces as a check (a few millimetres for a correct pose convention). The pose logic follows graspnetAPI's loadSceneModel / generate_scene_model as used in GraspFormer.

Not included: the grasp labels (grasp_label/, per object) and the per-scene collision labels (collision_label/, 184 MB for this scene); both are in the original GraspNet-1Billion release.