MultiCamRobolab

A real-world multi-camera RGB-D dataset of dynamic scenes, recorded in a laboratory with two Microsoft Azure Kinect cameras. Ground-truth camera poses come from a Qualisys motion-capture system, and ground-truth depth comes from the Kinect depth sensor.

Code: multiple-view-dynamic-reconstruction

Scenes

19 sequences in 4 scenarios:

Scenario	Directory	Sequences	Description
Robodog (overlap)	robodog_overlap/	robodog_01–04	A quadruped robot walks; the two cameras share enough view overlap
Robodog (non-overlap)	robodog_nonoverlap/	robodog_05–08	Same as above, but the cameras have little or no overlap
RoboArm	roboarm/	roboarm_01–04	A 6-DoF manipulator performs digging motions
DynamicHuman	dynamichuman/	dynamichuman_1, 4, 5, 6, 7, 8, 9	A person walks or works in the scene

Each sequence has 150 frames per camera (roboarm_04: 126) at 512×384. The frame rate differs between sequences (about 10–19 fps), so use the timestamps instead of assuming a fixed rate.

Three-camera extension

threecam/ holds 3 additional sequences recorded later with three Azure Kinect cameras (cam01–cam03). They are not part of the 19-sequence benchmark used in the paper, and they have no dynamic masks.

Sequence	Frames per camera	Description
robodog4_3cameras	70	A quadruped robot walks
robodoghuman2_3cameras	70	A quadruped robot and a person move in the scene
robodoghuman3_3cameras	90	A quadruped robot and a person move in the scene

Layout

MultiCamRobolab/
├── calib/                         # intrinsics of the 2-camera benchmark sequences
│   ├── cam01_intrinsics.txt       # 512×384, fx fy cx cy
│   ├── cam01_raw_1280x720.txt     # original sensor intrinsics + distortion (see "Data processing")
│   ├── cam02_intrinsics.txt
│   └── cam02_raw_1280x720.txt
├── threecam/
│   ├── calib/                     # same files for the 3-camera sequences (cam01–cam03)
│   └── <sequence>/                # same layout as below, with cam03 and without masks/
└── <scenario>/<sequence>/
    ├── cam01/                     # RGB, rgb_<sec>_<nsec>.png
    ├── cam02/
    ├── cam01_depth/               # GT depth, depth_<sec>_<nsec>.png (uint16, millimeters)
    ├── cam02_depth/
    ├── masks/cam01/, masks/cam02/ # dynamic-object masks, same file names as RGB (uint8, 0 / 255)
    ├── gt_poses/cam01_poses.txt, cam02_poses.txt
    └── intrinsics/cam01_intrinsics.txt, cam02_intrinsics.txt

File formats

• RGB: 8-bit PNG, 512×384.
• Depth: 16-bit PNG in millimeters; 0 means invalid. This is lossless with respect to the Kinect's native millimeter resolution. Depth timestamps differ from RGB timestamps by under 1 ms, so frames correspond by sorted order. The depth is not undistorted; see Data processing.

  depth_m = cv2.imread(path, cv2.IMREAD_UNCHANGED).astype(np.float32) / 1000.0
  

• Masks: 8-bit PNG, 255 = dynamic object, 0 = static background. They were generated with SAM 3, and the dynamic objects were selected manually.
• GT poses: TUM format, one line per frame and no header: timestamp tx ty tz qx qy qz qw (camera-to-world, quaternion in x y z w order, meters).
• Intrinsics: one line fx fy cx cy (pixels), for the 512×384 images. Every camera has its own intrinsics, and each sequence's intrinsics/ folder is authoritative. They are constant within each recording session, so calib/ (benchmark) and threecam/calib/ (3-camera) are given for convenience. Note that cam02 differs between the two sessions.

Data processing

Recording. Each Azure Kinect was recorded with its ROS driver: color from /kinect_X/rgb/image_raw (1280×720, 30 fps) and depth from /kinect_X/depth_to_rgb/image_raw (registered to the color camera, uint16 mm). Camera poses come from Qualisys rigid bodies mounted on each Kinect. RGB, depth and mocap messages are matched with an approximate time synchronizer, and frames with invalid (NaN) mocap poses are dropped. The original intrinsics and distortion from /kinect_X/rgb/camera_info are provided as calib/camXX_raw_1280x720.txt: fx fy cx cy k1 k2 p1 p2 k3 k4 k5 k6 (OpenCV rational model).

RGB. Each frame is processed as:

1280×720 (raw) ─ cv2.undistort(K, dist) ─ resize ×384/720 (bilinear) → 682×384 ─ center crop → 512×384

The center crop removes 85 px on the left and right, so the images cover the central 960×720 region of the original frame. cv2.undistort keeps the original K, so the RGB images are undistorted pinhole images.

Depth. Depth is converted to meters, resized with nearest-neighbor interpolation and center-cropped exactly like RGB, but it is not undistorted. Because the registered depth follows the distorted color geometry, it is slightly misaligned with the undistorted RGB: about 4 px (median) at 512×384, growing to about 15 px at the image corners and under about 1 px near the center. All results in our paper use the depth as released. If you need pixel-accurate alignment, undistort the depth with the raw calibration before cropping (nearest-neighbor remap).

Intrinsics. fx, fy, cx, cy are the raw values scaled by 384/720, with cx shifted by the 85 px crop. (The image is resized to int(1280·384/720) = 682 px wide, so the true horizontal scale is 682/1280; fx and cx are therefore about 0.1% / 0.3 px off, which is negligible.)

GT poses. The rigid-body pose of each Kinect is converted to the camera optical frame (x right, y down, z forward): R_cam = R_body · R_y(+90°) · R_z(−90°), and a constant offset (0.0039, −0.032, −0.033) m is added to the translation in the world frame. The same constant applies to every frame and camera, so it only shifts the world origin and leaves all relative poses (between frames and between cameras) unchanged. The result is a camera-to-world pose in the mocap world frame, with the mocap timestamps. No separate rigid-body-to-camera (hand-eye) calibration was applied. In sequences with view overlap, reprojecting GT depth between cameras with the GT poses agrees to about 1–4 cm (median).

Clipping. Each sequence is a contiguous window of [start, end) synchronized frames cut from the full recording. The effective frame rate is set by synchronization, which is why it differs between sequences.

Download

huggingface-cli download shuooru/MultiCamRobolab --repo-type dataset --local-dir ./data/MultiCamRobolab

Citation

Coming soon.