SceneNN

RGB-D indoor scenes with instance-level mesh annotation, from SceneNN (Hua et al., 3DV 2016). This repo is the HKUST public release with frames extracted from the .oni videos.

83 Asus Xtion scenes. 29,441 synced RGB–depth pairs, kept at 2 fps from the ~30 fps capture.

Layout

main/{id}/
  {id}.ply                  reconstructed mesh
  {id}.xml                  instance tags and oriented boxes
  trajectory.log            original camera→world poses (full rate)
  rgb/{frame}.jpg           640×480 color. image/ is a symlink to rgb/
  depth/{frame}.png         640×480 16-bit depth, millimeters
  depth_vis/{frame}.png     turbo colormap of depth (viewing only)
  timestamp.txt             seq, color frame id, depth frame id, timestamps (µs)
  intrinsic.txt             3×3 K
  camera_pose/{frame}.txt   4×4 camera→world for that rgb frame
  instance-segmentation/    mesh labels projected into the camera
oni/{id}.oni                raw OpenNI RGB-D video
intrinsic/asus.ini          source of K (fx = fy = 544.47329, cx = 320, cy = 240)
images/{id}/                official preview renders, not the RGB-D frames
contrib/color_ply_fixed/    {id}.ply color meshes, same vertex layout as main/{id}/{id}.ply

{frame} is the OpenNI color frame id, 6 digits (000001.jpg). pose_check/ and main/005/web* are projection checks and a mesh viewer.

Frames

Color and depth are paired with the SceneNN playback sync: timestamps within 33 ms, then subsampled to 2 fps. timestamp.txt starts with

# seq color_fid depth_fid color_ts_us depth_ts_us

seq is the 1-based index of the synced pair. trajectory.log is 0-based over those pairs, so the pose for a row is trajectory.log entry seq - 1, written to camera_pose/{color_fid}.txt. A frame is omitted there when that entry is missing (29,117 poses for 29,441 rgb frames).

Each block in trajectory.log is a frame index line plus a 4×4 matrix. The matrix maps a point from camera space to world space. intrinsic.txt is row-major K.

Mesh labels

{id}.ply is binary little-endian. Vertices are x y z, nx ny nz, uint label, uchar red green blue. label is the instance id. 0 is not used.

{id}.xml uses the same ids:

<label id="610041" color="195 224 75" text="bed" note="" area="100035"
       obbox="cx cy cz dx dy dz qx qy qz qw"/>

text is the tag and is often empty. obbox is center, size along x y z, then quaternion x y z w. Boxes come from eigen decomposition and are only tight on elongated objects.

instance-segmentation/{frame}.npy is uint32 of shape (480, 640). The value is the mesh instance id. 0 means the ray missed the mesh. {frame}.png is the instance color for viewing. Masks are a raycast of {id}.ply through camera_pose and intrinsic.txt (29,045 frames). A face whose vertices disagree takes the vertex with the largest barycentric weight.

License

Free for education and research. Contact scenenn[at]gmail.com before commercial use. Cite the original paper if you use the data:

@inproceedings{scenenn-3dv16,
    author = {Binh-Son Hua and Quang-Hieu Pham and Duc Thanh Nguyen and Minh-Khoi Tran and Lap-Fai Yu and Sai-Kit Yeung},
    title = {SceneNN: A Scene Meshes Dataset with aNNotations},
    booktitle = {International Conference on 3D Vision (3DV)},
    year = {2016}
}