3DLF-Scan dataset: multi-sensor 3D light-field and structured-light scans of 3D-printed Stanford figures

The 3DLF-Scan dataset contains multi-sensor 3D acquisitions of seven 3D-printed canonical Stanford models: bunny, dragon, asian_dragon, armadillo, happy, lucy, thai_statue.
Each figure was printed in dark matte resin and captured on a motorised turntable with two PhotonicSense light-field cameras (apiCAM PRO and apiCAM CUBE) and a Revopoint Miraco structured-light scanner, covering a full 360° sweep per object.

For each light-field camera and object, the dataset provides per-view:
– RGB images (rgb/) and optional brightened RGB (rgb_bright/)
– Dense depth exported from the camera (depth/)
– Sparse or high-confidence depth visualizations (depth_sparse/)
– Sparse depth values as NumPy arrays in millimetres (depth_npy/)
– High-fidelity dense depth completion maps (depth_hifi/) and their colour visualizations (depth_hifi_vis/)
– Binary masks in RGB and depth space (masks_rgb/, masks_depth/)
– Per-view point clouds and a poses_metric.json file with 4×4 camera-from-object poses and turntable geometry (pcd/).

For each object, additional sequences from the Revopoint Miraco scanner provide RGB images, metric depth maps, depth visualizations, and per-view point clouds derived from the vendor cache files.

The dataset also includes camera calibration folders (calib_pro/, calib_cube/) with chessboard images, intrinsic parameters (rgb_optic.json), and undistortion maps for OpenCV-style remapping.

3DLF-Scan is intended for developing and benchmarking methods for 3D reconstruction, light-field depth estimation, sparse-to-dense depth completion, multi-view fusion (e.g. Gaussian splatting, TSDF), cross-sensor registration, and pose-aware learning on controlled tabletop scenes.