Deterministic ground truth from video.
Tattva3D converts monocular video and known scene constraints into auditable camera poses, trajectories, and speed estimates for forensic analysis and liability determination.
Working components of the current analysis pipeline.
Lens Validation
Intrinsics and distortion are solved from measured 2D ↔ 3D correspondences and checked by reprojection residuals on held-out points before any downstream estimate is used.
Camera Trajectory Recovery
Per-frame pose is recovered from tracked image points and known scene geometry. Each frame's solution is independent and inspectable on its own merits.
Constrained Motion Analysis
Velocities and trajectories are derived in scene-aligned units from constrained motion, with explicit uncertainty — never a single unqualified number.
Reviewable Runs
Inputs, parameters, and intermediates are logged so a run can be reproduced and challenged step by step from the audit manifest.
Camera-to-scene alignment from measured correspondences.
The foundation underneath every downstream estimate: a calibrated lens and a set of measured 2D ↔ 3D correspondences are used to lock the camera into scene-aligned coordinates. Reprojection residuals on held-out points keep the alignment honest.
Everything that follows — per-frame pose, constrained motion, speed — is derived from this anchored frame of reference rather than from inferred geometry.
Scene-constrained human motion from monocular video.
An early prototype recovering an individual's motion from a single video and re-expressing it inside the same constrained scene geometry used for camera and vehicle estimates. The human figure is anchored to measured ground and the recovered camera — not free-floating.
Note the vehicles in the background: they jitter frame-to-frame because, in this clip, they are tracked without a metric anchor of their own. That instability is the exact failure mode the next prototype addresses — by constraining each vehicle to scene geometry and known dimensions, rather than letting it drift.
Exploratory R&D, not a shipped capability. Shown to illustrate the direction: extending the same scene-constrained, inspectable approach from vehicles to people.
Footprint recovery for subject path reconstruction.
Building on the recovered human motion, per-step footfall contacts are extracted and projected onto the scene-aligned ground plane. The result is an explicit subject path — a sequence of measured ground contacts in scene coordinates — rather than a free trajectory floating in space.
Because the path is derived from constrained motion against known geometry, it reconstructs on surfaces that leave no physical trace: asphalt, tile, concrete, indoor flooring. The footprints exist in the analysis even when they never existed at the scene.
Exploratory R&D — shown to illustrate how scene-constrained motion extends into defensible path evidence for pedestrian and bystander analysis.
Semantic vehicle masking for constrained speed estimation.
The direct response to the background jitter seen in the previous clip. Here, the vehicle is segmented per frame and locked into the same scene-aligned 3D environment, with the vehicle's known physical dimensions acting as a metric anchor. The bounding box stays dimensionally consistent across frames instead of drifting.
That dimensional anchoring is what turns per-frame masks into a defensible speed reading: motion is derived from constrained, measurable inputs rather than an unconstrained track.
Exploratory R&D — not a validated speed measurement. Shown to make the working approach visible.
Dimensionally anchored vehicular meshes.
The background jitter seen in earlier clips — vehicles drifting frame-to-frame because they lacked a metric anchor — is resolved here by constraining the vehicle mesh to its known physical dimensions and locking it into scene-aligned coordinates.
Rather than tracking an unconstrained bounding box, the mesh itself is dimensionally anchored: every vertex is held to measurable geometry, so the vehicle stays fixed in scale and position relative to the ground plane across the entire sequence.
This is the pieces coming together. Scene calibration, constrained segmentation, and metric anchoring combine into a single inspectable output: a vehicle whose motion is derived from geometry it cannot violate. Exploratory R&D — shown to illustrate the full pipeline from evidence to measurement.
Many video-to-3D systems infer geometry the camera never observed. In a forensic context, an inferred surface is not evidence — it is an assumption presented as a measurement.
Tattva3D works the other way. Every estimate is anchored to a calibrated lens, tracked image points, and scene geometry that can be measured or surveyed. Nothing downstream is computed from unseen structure, and every intermediate remains open to inspection.
The output is not a guessed scene. It is an inspectable chain from evidence to measurement.
Forensic engineers
Recover scene-aligned measurements from video evidence with a defensible computational chain.
Accident reconstruction teams
Pair video with survey or scan data to constrain trajectories and speeds without proprietary scene reconstruction.
Insurers and liability teams
Inspect, reproduce, and challenge measurements without relying on a black-box estimate.
- —Not a generative video-to-3D toy
- —Not a photogrammetry replacement
- —Not a black-box “AI says the speed was X” system
- —Not optimized for visuals before mathematical validity
Work today centers on constrained motion recovery from monocular video. Each component below is held to the same requirement: the result must be reproducible from a logged manifest by an outside reviewer.
- F.01Lens calibration
- F.02Tracked points
- F.03Per-frame camera pose recovery
- F.04Constrained motion analysis
- F.05Speed estimation
- F.06Auditability and rerunability