iPhone LiDAR 360° RGB-D Object Scans (pilot): four household objects with per-frame poses, confidence maps and verified metric scale

A pilot release of 4 everyday objects captured as posed RGB-D sequences with an unmodified iPhone 13 Pro (LiDAR) using the Stray Scanner app. Each object is a complete 360° handheld orbit with per-frame camera poses, per-frame intrinsics, LiDAR depth and per-pixel confidence maps. 1,019 frames, 1.07 GB. The emphasis is on verification rather than volume. Every object's reconstruction is checked against a physical measurement of one rigid dimension (errors span -5.0% to +2.7%); the metric scale of the depth is confirmed independently of any physical reference; the camera pose convention was established empirically rather than assumed; and the objects that failed to reconstruct are documented rather than quietly dropped. Contents per object: 1920×1440 colour frames (JPEG q92 4:4:4), 256×192 uint16 millimetre depth maps (lossless PNG), 256×192 ARKit confidence maps, per-frame 6-DoF poses in OpenCV camera convention, per-frame intrinsics, IMU, schema-validated metadata, TSDF-fused geometry, and SHA-256 checksums. Objects: air_purifier, supplement_container, dumbbell, glasses_case. Intended uses: per-scene reconstruction and novel-view synthesis (each object is a self-contained test scene with known poses and verified metric scale); validating RGB-D loaders and ingest pipelines; and depth super-resolution and completion research, where each frame pairs a high-resolution colour image with a low-resolution depth map and confidence mask from the same optical axis. Not a training set. Four objects and roughly one thousand frames is far short of what any data-driven model requires. The set also cannot support category-level claims: it contains four unrelated categories with a single instance each. The published objects are a selected sample — seven were captured and three excluded for poor reconstruction — so any accuracy figure computed on it is an optimistic bound rather than a typical case. The accompanying datasheet states this in full. Known limitations include a practical lower bound of roughly 15 cm on object size at these working distances, degraded depth on specular, transparent and very dark surfaces, and edge rounding of about 0.5–1 cm from ARKit's depth smoothing. Fused meshes are provided for quality assessment and are not ground truth. This is version 0.1.0, a pilot. The intended growth path is roughly 10–15 objects with 2–3 instances per category.