Interactive real Camera–LiDAR projection

The browser's Camera–LiDAR tab projects real KITTI raw Velodyne returns onto the corresponding rectified image_02 photograph. Select frame 0060 or 0090 from development drive 2011_09_26_drive_0005_sync. No rendered sensor data is used.

Move roll, pitch and yaw (±5°), or expand the translation controls (±0.20 m). Inject +3° pitch and Restore vendor offer reproducible comparisons. The divider shows the supplied KITTI calibration on the left and the manually adjusted projection on the right. Every control recomputes projections locally; it does not play a pre-rendered video or run a calibration optimizer.

The displayed pixel displacement compares the same points visible in both projections, before clipping either overlay to the divider. It measures how much the manual offsets move the projection, not calibration accuracy. There is no live pass/fail verdict. Sensor timing, occlusion and imperfect calibration may leave mismatches even at the supplied vendor setting.

Sources and attribution

Credit: A. Geiger, P. Lenz, C. Stiller and R. Urtasun / KITTI. KITTI raw data and the derived photographs, sampled points and projection assets remain under CC BY-NC-SA 3.0. They require attribution and non-commercial use. Calibrex's source-code license does not relicense these data.

download-receipt.json records the official archive URLs, member names, byte counts and SHA-256 hashes. source/ preserves the original photographs and date-level calibration files. The two compressed point subsets preserve original Velodyne returns and their indices in the original scans.

Reproduce and inspect

Run PYTHONPATH=src python tools/build_camera_lidar_demo.py from the repository root. The generator uses the committed photographs, calibration files and point subsets; no network or ROS installation is needed. --prepare-from DIR can rebuild the subsets from the original downloaded scans and receipt.

Points are transformed by R_rect_00 × T_cam0_velo, then translated by K⁻¹ × P_rect_02[:,3] to enter rectified camera 2. This includes the projection matrix's camera baseline; omitting it would misalign the overlay. Manual offsets apply Rz(yaw) × Ry(pitch) × Rx(roll) to these camera points, followed by XYZ translation in metres. Intrinsics and photograph resolution stay fixed.

For display, returns have camera depth 3–60 m, original Velodyne Z above −1.6 m, and a generous projected margin around the photograph. Up to 6,000 points are sampled evenly by original scan index. Colours represent camera Z, clipped to 5–40 m. This is a display subset, not an input to an accuracy evaluation.

The generated index binds original and derived inputs, generator, images and point assets to SHA-256. Per-frame assets also contain independently generated SciPy projection references for JavaScript verification. Images and point assets load only when their scene is selected.

The earlier recorded Camera–LiDAR check GIF remains available separately; its saved endpoint verdicts are not applied to arbitrary slider positions. Broader board/edge/mutual-information method comparison remains a separate demo task.