PhysLens

PhysLens: video in, measured physics out. A video shows what happened, but rarely how fast. Crash investigators, physics students and labs still read speeds and distances off footage by hand. PhysLens turns ordinary video (or one photo) into measured 3D points, follows each object as one rigid body over time, finds the physical events (contact, crush, rebound, turn) and explains them. Every number is checked against an official measurement, and what fails is written down too. Checked against official data (public footage): - NHTSA 9070, SUV into a rigid wall (1,000 fps): approach 56.81 km/h vs 56.55 km/h speed trap; crush 0.728 m at 75 ms vs 0.707 m at 74.1 ms; crush curve within 1.5 cm of the accelerometers. - NHTSA 15756, car sideways into a pole: 31.29 km/h at 74.6° vs 31.09 km/h at 75°. - NHTSA 8555, pickup into a standing SUV: approach 69.09 vs 69.1 km/h; SUV speed change +47.4 vs +47.8 km/h. - NHTSA 10072, moving barrier at 15°: 90.72 vs 90.33 km/h. - Box truck under the 11foot8 bridge: impact at frame 196.6 (hand label 196-200), clearance 3.63 m vs the 3.556 m sign, truck flagged too tall. - Classroom ball drop: g = 9.73 ± 0.21 m/s²; the next bounce is predicted and locked with a hash before the frames that settle it. - Multi-camera: a football pass in CMU Panoptic Studio, 31 calibrated cameras: the ball triangulated from 4-6 cameras per moment; its flight gives g = 9.49 ± 0.08 m/s² with air drag; the people as solid bodies (space carving, checked by a held-out camera and by stereo to about 2 cm). What you can use: - Lab: drop and bounce, crashes and more, from samples, records or your own clip; predictions locked before reality settles them; what-if replays (gravity, bounciness, speed). - 3D simulator: walk around each measured world, scrub through the event, see which points are measured and which are predicted. - Scan: one photo becomes named parts and measurements. - Ask by voice or text: answers come only from the measurements. - Specs agent: identifies the object (e.g. by VIN) and looks up its published dimensions with sources, cross-checked between sources and against the video. Runs on a laptop; the paid Meta calls are capped and logged.