Research Note
Point Tracking Deployment Review Record
Freeze the exact repository revision, checkpoint hash, dependency lock, model mode, preprocessing, configuration, and evaluation record. Record code, model, dataset, medi
Point Tracking Deployment Review Record
Artifact boundary
Freeze the exact repository revision, checkpoint hash, dependency lock, model mode, preprocessing, configuration, and evaluation record. Record code, model, dataset, media, and dependency licenses separately.
The official CoTracker repository says most of the project uses a CC BY-NC license, with some components under other terms. This is a commercial-use and provenance gate, not a footnote.
Pipeline review
| Layer | Evidence |
|---|---|
| Capture | Camera, source, consent, rights, time, location |
| Ingestion | Codec, validation, malware boundary, dropped frames |
| Preprocessing | Decode, resize, crop, color, frame rate, normalization |
| Inference | Checkpoint, mode, queries, windows, hardware, batching |
| Output | Coordinate system, visibility, confidence, schema, consumers |
| Storage | Raw video, derived tracks, retention, encryption, access |
| Monitoring | Latency, memory, queue, drift, sampled failures, incidents |
| Control | Human review, abstention, stop condition, rollback, deletion |
Runtime
Measure end-to-end capture-to-consumer latency, not only model time. Test warm and cold starts, different query counts, resolution, long streams, queue growth, device contention, corrupted input, dropped frames, and hardware loss.
Online mode is not automatically real-time. Offline mode may use future frames and may be unsuitable for a causal decision. The selected mode must match the workflow.
Silent failure
A crashed job is visible. A plausible but wrong trajectory can pass downstream unnoticed. Monitoring should sample rendered tracks, compare with human-reviewed clips, detect impossible jumps or stale coordinates, and measure the consequence in the consuming workflow.
Release stages
Start with offline replay, then shadow mode, then a bounded assisted workflow, then a narrow production release. Each stage requires an owner, threshold, human override, incident path, and exercised rollback.
Governance boundary
The deployment record must state purpose, affected people, access, data use, prohibited uses, monitoring, complaints, and reapproval triggers. A new camera, population, purpose, market, or decision consequence requires renewed review.
This framework is not authorization for surveillance, autonomous control, medical use, employment monitoring, or another high-stakes application.
Sources
Follow the evidence.
- youtu.be: BNTcjZ0Ym38youtu.be
- ai.meta.com: sam2ai.meta.com
- proceedings.neurips.cc: 58168e8a92994655d6da3939e7cc0918 Abstract Datasets and Benchmarksproceedings.neurips.cc
- arxiv.org: 2410arxiv.org
- open.spotify.com: 26JgnnwjvK5vYIdRofV8ntopen.spotify.com
- github.com: co trackergithub.com
- cotracker3.github.iocotracker3.github.io
- NIST AI Risk Management Frameworknist.gov
- vggsfm.github.iovggsfm.github.io
- daltonanderson.ghost.io: metas cotracker 3 a leap in ai object trackingdaltonanderson.ghost.io
- arxiv.org: 1803arxiv.org
- ecva.net: 3526 ECCV 2020 paperecva.net
- github.com: tapnetgithub.com
- raw.githubusercontent.com: LICENSEraw.githubusercontent.com
- tapvid.github.iotapvid.github.io
- NIST Privacy Frameworknist.gov
- arxiv.org: 1504arxiv.org
- openaccess.thecvf.com: Karaev CoTracker3 Simpler and Better Point Tracking by Pseudo Labelling Real Videos ICCV 2025 paperopenaccess.thecvf.com