Skip to content

Add causal image-motion alpha with measured forecast replay - #7

Open
Tarzelf wants to merge 2 commits into
feat/observation-pilot-v1from
feat/motion-observation-v0.2
Open

Tarzelf wants to merge 2 commits into
feat/observation-pilot-v1from
feat/motion-observation-v0.2

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Oct 11, 2026 •

Copy link
Copy Markdown
Contributor

When reviewing motion footage, DVIDIA can now follow a manually selected image region and show observed image speed alongside causal forecasts at 100, 250 and 500 ms. The new dvidia-motion command analyzes permitted local footage, validates receipts, runs authored fixtures and serves a branded read-only replay. It uses CPU sparse feature tracking and actual integer video timestamps; it trains no model and downloads no weights.

Tracking losses withhold forecasts. The replay uses the latest observation at or before the playhead and expires stale predictions. Source, seed, clock, protocol, camera display geometry and causal rows are checked before replay; benchmark scores are recomputed against hash-bound evaluator references. Future references never enter tracking.

The first measured batch at commit ade2b541d8ca4a785bbd1effdfd33eb2a306f150 preserves conditional errors, answer coverage, abstentions, timing scope and failure tails. All 720 frames were attempted. Processing took 1.454 seconds for 24 seconds of authored video on one ARM64 machine; this is offline throughput, not live latency. Occlusion and low texture lose tracking. Direction reversal reaches approximately 90 pixels worst forecast error at 500 ms. Speeds and errors remain image-plane quantities; physical motion, independent real-scene accuracy, live robot interaction and calibrated uncertainty remain unqualified.

Intake and replay reject nonzero rotation, nonidentity display matrices and explicit non-square pixel aspect ratios. A bounded metadata probe has a 64 KiB output cap and a 10-second owned-process watchdog. Older identity receipts remain readable; consistently rehashed rotated media cannot bypass the replay guard.

Includes source installation instructions, a full v0.2 plan and updated execution map. Original motion code/fixtures use MIT; existing upstream Apache notices remain packaged. The published v0.1 tag and Hugging Face review demo are unchanged.

Validation:

  • 389 local tests passed on the final source, including actual FFmpeg decoding, real rotated/anamorphic metadata fixtures and 25 replay integrity/HTTP tests. The first measured batch remains frozen; its processing numbers predate the added geometry probe.
  • Six browser checks passed in installed WebKit: causal playhead, controls, exact receipt export, loss abstention, desktop/mobile layout and expiration.
  • Updated installable wheel built, inspected and installed in an isolated environment: motion entry point, replay HTML and three license files verified; old authored receipts pass inspection. No footage or model weights bundled.
  • CI adds dedicated CPU motion tests plus authored benchmark/inspection, and base-platform command/core/scorer checks.

Measured evidence: benchmark. Source installation: motion alpha.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant