luohwu/SpineDepth_perframe

SpineDepth per-frame: raw RGB-D frames, CT meshes and per-frame ground-truth poses

This is a frame-wise release of the SpineDepth dataset (Liebmann et al., 2021) for CT to intraoperative RGB-D registration and vertebra pose estimation. It contains the raw per-frame data: RGB-D point clouds with a vertebra label on every point, the CT bone meshes, and the optical-tracking pose of every vertebra in every frame. A small script turns the raw data into per-vertebra point clouds in camera space and in CT space.

How this differs from SpineDepth_segmented

Our earlier release, luohwu/SpineDepth_segmented, contains merged point clouds. All tracked frames of a camera were fused into one downsampled CT-space cloud per vertebra, so each vertebra has two clouds, one per camera. It is a static registration set with no per-frame information.

This release keeps every RGB-D frame as its own point cloud. The frames are consecutive frames of the original 30 Hz recordings. Each frame comes with the full raw scene in camera space, a vertebra label per point, and the pose of every vertebra in that frame. Nothing is fused, filtered or downsampled. Use SpineDepth_segmented for clean CT to RGB-D surface registration. Use this dataset when you need single depth frames, short temporal sequences, per-frame poses, the surrounding scene, or the raw sensor noise.

Fully reconstructed per-frame files are large, so we ship the compact raw data together with the reconstruction script instead of prebuilt point clouds. Reconstructing one specimen takes a few minutes on a desktop machine.

Quick start

You need Python 3.8 or newer with numpy and open3d 0.15 or newer. The script finds raw/ and writes reconstructed/ relative to its own location, so the folder can be unpacked anywhere and no paths need editing. Run it from any working directory.

pip install -r requirements.txt
python reconstruct_frames.py                  # reconstructs specimen 2 (default) into ./reconstructed
python reconstruct_frames.py --specimen 7     # one specimen (1..10)
python reconstruct_frames.py --specimen all   # all ten specimens (~35 GB, ~20 min with 12 workers)
python reconstruct_frames.py --specimen 3 --workers 8 --out-root /some/other/folder

Dataset summary

• 10 cadaveric specimens (IDs 1 to 10), lumbar levels L1 to L5.
• 50 CT bone meshes (.stl), one per specimen and level.
• 2 RGB-D cameras (Stereolabs ZED Mini, 1920×1080, 30 Hz) per recording, 40 recordings per specimen (10 pedicle screws × 4 surgical steps).
• Per frame: an RGB-D point cloud in camera space with a vertebra label on every point, and the pose of each of the five vertebrae (CT to camera).
• Frames per recording and camera: 4 consecutive frames (frame_0 to frame_3, the first four frames of each video). A few recording and camera pairs have fewer or no frames; see the per-specimen overview below. The full videos, several hundred frames per recording, are available as SVO files from the original SpineDepth release.

Directory structure

SpineDepth_perframe/
├── README.md
├── requirements.txt
├── reconstruct_frames.py                      # builds ./reconstructed from ./raw
├── raw/                                       # shipped
│   └── Specimen_<s>/                          # s ∈ {1..10}
│       ├── STL/
│       │   └── L1.stl  L2.stl  L3.stl  L4.stl  L5.stl       # CT bone meshes, each in its own CT space
│       ├── transformation_matrices.txt                        # unused by the script, see "Raw file formats"
│       └── recordings/
│           └── recording<r>/                  # r ∈ {0..39}
│               ├── Poses_0.txt                # ground-truth poses for camera 0, every video frame
│               ├── Poses_1.txt                # ground-truth poses for camera 1, every video frame
│               └── Specimen_<s>_recording_<r>_cam_<c>_frame_<f>.npz   # c ∈ {0,1}, f ∈ {0..3}
└── reconstructed/                             # created by reconstruct_frames.py (not shipped)
    ├── index.csv                              # one row per (specimen, camera, recording, frame, level)
    └── Specimen_<s>/
        ├── index.csv
        ├── CT/L1.stl … L5.stl                 # copied from raw
        ├── camera_0/
        │   └── recording_<rr>/                # rr ∈ {00..39}
        │       └── frame_<f>/
        │           ├── frame_camera_space.ply          # full raw RGB-D frame, camera space, RGB
        │           ├── frame_camera_space_labels.npy   # int8 label per point of the file above
        │           ├── poses_ct_to_camera.txt          # 5 rows (L1..L5), p_camera = T @ p_ct
        │           ├── L1_camera_space.ply  …  L5_camera_space.ply   # labelled points, camera space, RGB
        │           └── L1_ct_space.ply      …  L5_ct_space.ply       # same points in the CT space of that level
        └── camera_1/
            └── (identical layout)

Raw file formats (raw/)

File	Format	Content
Specimen_<s>_recording_<r>_cam_<c>_frame_<f>.npz	NumPy .npz, key arr_0, float64 array of shape (N, 7)	Columns x, y, z (millimetres, camera space), r, g, b (in [0, 1]), label (0 background, 1 to 5 = L1 to L5). One file per video frame f of camera c.
Poses_<c>.txt	text, 16 comma-separated floats per line	Poses for every video frame of camera c, 5 lines per frame in the order L1..L5. Line 5*f + (k-1) is the 4×4 matrix T of vertebra Lk at frame f, flattened row-major, with p_camera = T @ p_ct. The file covers the whole video, so it has many more frames than there are npz files.
STL/L<k>.stl	binary STL	CT-segmented bone surface of vertebra Lk, in its own CT space.
transformation_matrices.txt	text, 4 lines × 16 floats	Four 4×4 matrices carried over verbatim from the original preprocessing. The reconstruction script does not use them and their meaning is not documented here.

Reconstructed file formats (reconstructed/)

File	Format	Notes
frame_camera_space.ply	binary PLY, double x y z, uchar red green blue	Every point of the npz frame: bone, soft tissue, instruments, hands, background. Camera space.
frame_camera_space_labels.npy	NumPy int8, shape (N,)	Same point order as the ply above. 0 = background, 1 to 5 = L1 to L5.
poses_ct_to_camera.txt	5 lines × 16 comma-separated floats	The five pose rows of this frame, copied from Poses_<c>.txt. Line k = vertebra Lk.
L<k>_camera_space.ply	binary PLY, as above	Points of the frame whose label equals k. Camera space.
L<k>_ct_space.ply	binary PLY, as above	L<k>_camera_space.ply transformed by inv(T_k). Overlays directly on CT/L<k>.stl.
CT/L<k>.stl	binary STL	Copy of the raw mesh, stored once per specimen.
index.csv	CSV	Columns: specimen, camera, recording, frame, level, num_points_level, num_points_frame, camera_space_ply, ct_space_ply, ct_mesh. Paths relative to reconstructed/.

Coordinate spaces

Space	Definition	Files
camera space	Optical frame of the ZED Mini that captured the frame. Millimetres. Each camera and each recording has its own camera space, because the cameras were repositioned before every pedicle screw.	npz x y z, frame_camera_space.ply, L<k>_camera_space.ply
CT space of vertebra L	Coordinate frame of STL/L<k>.stl. Millimetres. Each vertebra has its own CT space. The five CT spaces of one specimen are not aligned with each other, because every vertebra was tracked by its own optical marker.	L<k>.stl, L<k>_ct_space.ply

For vertebra L in a given frame, p_camera = T_k @ p_ct, where T_k is the pose line of that vertebra (4×4 homogeneous). Equivalently L<k>_ct_space = inv(T_k) @ L<k>_camera_space.

Provenance

The data start from the original SpineDepth recordings (stereo RGB-D video tracked by an optical tracking system) and the CT bone models.

1. Frame export. The first four frames of each recording and camera were exported from the SVO videos as RGB-D point clouds in camera space. These are the npz files.
2. Poses. For every video frame and every vertebra, the pose T_k (CT to camera) comes from the optical tracker. Each vertebra carried its own marker, fixed with K-wires and calibrated to the CT in the original SpineDepth work. The poses are stored in Poses_<c>.txt.
3. Vertebra labels. The label column was derived from the tracked CT models; no model predicts it. The labels behave like a projected silhouette mask: every point along a camera pixel that falls inside the rendered outline of vertebra Lk receives label k. So points of instruments, hands or soft tissue in front of the bone, and depth outliers, can carry a bone label. Typical distances between labelled points and the CT surface are a few millimetres, with a tail up to about 2 cm. The script applies no 3D distance filtering, so you can choose your own cleaning step, for example keeping only points within 5 mm of the CT surface.
4. Reconstruction. reconstruct_frames.py splits each npz by label, writes the per-vertebra clouds in camera space, and transforms them by inv(T_k) into CT space.

Recording protocol (from Liebmann et al., 2021)

Each specimen received ten pedicle screws (L1 to L5, both sides). For every screw, four surgical steps were recorded separately: (1) entry point with a luer, (2) canal preparation with an awl, (3) canal integrity check with a K-wire, (4) screw placement. Both cameras were repositioned and recalibrated before every screw, so recordings 4n to 4n+3 share a camera pose and the camera pose jumps between such groups. Within a specimen the vertebrae barely move relative to each other. For specimen 2, the relative poses of adjacent levels vary by less than 0.3 mm and 0.2° over the whole procedure. The absolute pose drifts by up to a few millimetres within a recording as the surgeon works on the bone. The per-frame poses account for all of this.

Per-specimen overview

Specimen	Exposure (orig. paper)	Screw order (orig. paper)	Recordings with frames	Recording/camera pairs (of 80)	npz frames	Video frames per camera
1	Midline approach, soft tissue intact	L1–L5	4	4	16	11,278
2	Full exposure	L1–L3, L4–L5	40	80	320	18,194
3	Full exposure	L1–L2, L3–L5 (LLL, RRR)	40	79	313	16,747
4	Full exposure	L1–L5	40	79	307	18,746
5	Full exposure	L1–L3, L4–L5	40	80	320	15,649
6	Full exposure	L3–L4, L4–L5, L1–L2	40	68	272	15,126
7	Full exposure	L4–L5, L1–L3	40	80	320	12,871
8	Full exposure	L2–L4, L4–L5, L1–L2	40	80	320	13,844
9	Full exposure	L1–L5	40	74	276	12,202
10	Full exposure	L4–L5, L1–L3	30	55	220	12,825
Total					2,684	147,482

"Recording/camera pairs" counts the recording and camera combinations that have at least one npz frame; a complete specimen has 40 × 2 = 80. Most pairs have 4 frames; a handful in specimens 3, 4 and 9 have 1 to 3. "Video frames per camera" is the number of frames in the full original videos (the length of Poses_<c>.txt divided by 5). Screw-order groups are separated by commas. Previously inserted screws were removed before the next group unless they reappear in it. Specimen 1 was instrumented through a standard midline approach with the soft tissue left intact. Its bone surface is largely occluded and only four recordings were exported.

Reconstruction output is about 12.5 MB per frame, roughly 4 GB per full specimen and about 34 GB for all ten.

Usage

import numpy as np, open3d as o3d

root  = "reconstructed/Specimen_2"
frame = f"{root}/camera_0/recording_00/frame_0"

# CT <-> RGB-D registration of L3 (already in a common frame):
ct    = o3d.io.read_triangle_mesh(f"{root}/CT/L3.stl")
rgbd  = o3d.io.read_point_cloud(f"{frame}/L3_ct_space.ply")

# Pose estimation: predict T from camera-space points, compare against ground truth
pts   = o3d.io.read_point_cloud(f"{frame}/L3_camera_space.ply")
T_gt  = np.loadtxt(f"{frame}/poses_ct_to_camera.txt", delimiter=",").reshape(5, 4, 4)[2]  # L3

# Place the CT model into the camera-space scene
ct_in_camera = o3d.io.read_triangle_mesh(f"{root}/CT/L3.stl").transform(T_gt)
scene  = o3d.io.read_point_cloud(f"{frame}/frame_camera_space.ply")
labels = np.load(f"{frame}/frame_camera_space_labels.npy")

# Or work directly on the raw npz without reconstructing:
raw = np.load("raw/Specimen_2/recordings/recording0/Specimen_2_recording_0_cam_0_frame_0.npz")["arr_0"]
xyz, rgb, label = raw[:, :3], raw[:, 3:6], raw[:, 6].astype(int)
poses = np.loadtxt("raw/Specimen_2/recordings/recording0/Poses_0.txt", delimiter=",").reshape(-1, 5, 4, 4)
T_L3_frame0 = poses[0, 2]

License

Released under CC BY-NC 4.0, following the original SpineDepth dataset.

Citation

If you use this dataset, please cite our work and the original SpineDepth paper:

@article{wu2026neuralbonereg,
  title   = {NeuralBoneReg: An instance-specific label-free point cloud-based method for multi-modal bone surface registration},
  author  = {Wu, Luohong and Seibold, Matthias and Cavalcanti, Nicola A. and Ao, Yunke and Flepp, Roman and Massalimova, Aidana and Calvet, Lilian and F{\"u}rnstahl, Philipp},
  journal = {Medical Image Analysis},
  year    = {2026},
  doi     = {10.1016/j.media.2026.104133}
}

@article{liebmann2021spinedepth,
  title   = {{SpineDepth}: A Multi-Modal Data Collection Approach for Automatic Labelling and Intraoperative Spinal Shape Reconstruction Based on {RGB-D} Data},
  author  = {Liebmann, Florentin and St{\"u}tz, Dominik and Suter, Daniel and Jecklin, Sascha and Snedeker, Jess G. and Farshad, Mazda and F{\"u}rnstahl, Philipp and Esfandiari, Hooman},
  journal = {Journal of Imaging},
  volume  = {7},
  number  = {9},
  pages   = {164},
  year    = {2021},
  doi     = {10.3390/jimaging7090164}
}