Flashkernel/realsense-multicam-tabletop

RealSense Multi-Camera Tabletop

Synchronized multi-view RGB + stereo IR captures of a tabletop scene from 4 Intel RealSense cameras, recorded with fixed camera positions. Includes FoundationStereo depth for two scenes and chessboard-derived extrinsics for merging the views into a single point cloud.

Layout

scene_000NN/
├── camera_poses.json          # extrinsics, only in calibration scenes (see table)
└── <camera_serial>/
    ├── rgb/00000.jpg ... 00119.jpg      # 640x480 RGB
    ├── left/00000.jpg  ... 00119.jpg    # 640x480 grayscale, left IR
    ├── right/00000.jpg ... 00119.jpg    # 640x480 grayscale, right IR
    ├── stereo_depth/00000.npy           # (480,640) float64, metres — FoundationStereo output
    ├── stereo_aligned_depth/00000.npy   # (480,640) uint16 — above, reprojected to the colour frame
    └── meta_info.json                   # per-camera intrinsics + depth_scale

Camera serials, identical across all scenes: 234322305266, 336222300744, 336522303601, 339522301222

Contents

Every scene has all 4 cameras × 120 frames of rgb, left, and right. Depth and poses are not uniform across scenes:

scene	rgb/left/right	stereo_depth	stereo_aligned_depth	camera_poses.json	role
scene_00001	120	120	120	—	object
scene_00002	120	—	—	✅	chessboard calibration
scene_00003	120	—	—	—	object
scene_00004	120	—	—	—	object
scene_00005	120	120	120	—	object
scene_00006	120	—	—	✅	chessboard calibration

Depth is a derived artifact: it is regenerated by running FoundationStereo on left/ + right/, so the four scenes without it can be filled in locally. Calibration scenes only ever need rgb/.

Conventions

meta_info.json — depth_intrinsics and color_intrinsics each carry a 9-element intrinsic_matrix in column-major order, i.e. [fx, 0, 0, 0, fy, 0, cx, cy, 1]. To use it:

import json, numpy as np
m = json.load(open("scene_00001/234322305266/meta_info.json"))
K = np.array(m["color_intrinsics"]["intrinsic_matrix"]).reshape(3, 3).T
depth_scale = m["depth_scale"]   # 0.001 -> uint16 units are millimetres

Depth units — stereo_depth/*.npy is float64 metres. stereo_aligned_depth/*.npy is uint16; multiply by depth_scale (0.001) to get metres. Zero means no return.

camera_poses.json — maps each camera serial to w2c and c2w, both 4×4 row-major homogeneous matrices, as produced by OpenCV chessboard pose estimation. c2w is the inverse of w2c:

poses = json.load(open("scene_00002/camera_poses.json"))
c2w = np.array(poses["234322305266"]["c2w"])   # camera -> world

Merging views

Poses come from a chessboard scene; geometry and colour come from an object scene. This is valid only because the cameras never moved between them — pair an object scene with a calibration scene captured in the same session, matching cameras by serial:

object scene_00001 + poses from scene_00002

Reproject each camera's stereo_aligned_depth with its color_intrinsics, transform by that camera's c2w, and concatenate to get a single merged cloud.

Provenance

Captured and processed with the realsense_multicam_tabletop pipeline (RealSense capture → FoundationStereo → depth-to-colour alignment → chessboard pose estimation → merge).