Repository navigation
Working local perception engine with 100 licensed internet images and a 50-second H.264 results film.
- Object detection, optional camera/video inputs, ByteTrack tracking, JSON outputs.
- Optional EasyOCR, YOLO26 pose, MediaPipe facial blendshapes, relative depth, and Cityscapes segmentation.
- 100 originals plus 600 controlled degradation tests; full optional perception on all originals.
- 80/20 fine-tuning split, 3 actual training epochs, measured AP and latency.
- 12 executed tests passed; source includes uv lockfile and CI.
Download 100-images-dataset.zip for originals, labels, and attribution; extract it in the repository root. 100-images-analysis.zip contains all full-pipeline visualizations and JSON. 700-degradation-results.zip contains original/combined views and JSON for every degradation experiment. The experimental fine-tuned checkpoint performs worse than the pretrained baseline, which remains the default. The film animates independent photographs; it does not claim tracking across them. No camera was accessed.
Code: AGPL-3.0. Images retain the license listed in ATTRIBUTION.md and manifest.json. Optional upstream models retain their own license terms. SegFormer is for research/evaluation under its upstream non-commercial terms. See reports/RESULTS.md for measurements, evaluation caveats, and unsupported temporal action/trajectory features.