GPU visual attention with YOLO, MediaPipe, a video API and evaluated releases.
161
GitHub source, demo and full documentation
YOLOv8n person detection runs with ONNX Runtime CUDA; MediaPipe head and eye landmarks run on CPU. One shared pipeline tracks session-local spatial IDs, estimates a visual heuristic score and produces JSON/JSONL and face-only H.264 annotations. An included web UI accepts video uploads and plays the result. No cloud account or API key is required.
Requires Linux x86-64 NVIDIA GPU containers. Tested native GPU: GTX 1660 Ti with 6 GiB VRAM. The complete image contains both production model bundles and pinned libraries; no Git clone or startup model download is needed.
docker pull bmaged23/multi-person-attention-tracking:gpu
docker run --rm --name attention --gpus all \
-p 127.0.0.1:8770:8770 \
-v attention-results:/app/data \
bmaged23/multi-person-attention-tracking:gpu
After initialization, open http://localhost:8770. Upload a clip, select its
processing duration, and download the summary, observations or annotated MP4.
The volume retains bounded results. Stop with Ctrl+C. For CPU mode of this same
bundle omit --gpus all and add -e ATTENTION_PERSON_DEVICE=cpu.
HTTP endpoints expose readiness/configuration and sequential upload jobs with
status, summary, observations and video artifacts. Open /docs for interactive
examples. Defaults limit uploads to 64 MiB, source duration to 10 seconds and video
resolution to 1080p. This is a local single-user service without authentication.
API guide · Docker guide · CI/CD guide
The release workflow tests the candidate image with 100 human-annotated COCO images, synthetic tracking trajectories, real API video jobs and browser upload/playback. Only passing candidates are published; public distribution is then tested through anonymous pull and actual GPU/CPU inference. Detailed measured results and release evidence are recorded in the evaluation report.
Native observations: detector median 0.401 s OpenCV CPU versus 0.070 s CUDA on 64 images, with identical detection counts. Two 50-frame AMI workflows took 48.9 s versus 14.8 s with identical observation JSONL. Component and workflow comparisons are distinct. These small exposed suites do not establish broad accuracy or real-time throughput.
Attention labels are head/eye visual heuristics, not comprehension measurements. IoU tracking can swap IDs at crossings and detection can miss occluded people. Source is AGPL-3.0-only; models and sample media retain their upstream terms. Notices and attribution.
Passing build, GPU gate, publication and public inference.
Immutable reference: bmaged23/multi-person-attention-tracking@sha256:bc953672a844f89de350924d680397b1d6cc659a48c26fafe8b094d176de765c.
The image passed 41 tests, 100 human-annotated detection cases, four synthetic tracking thresholds, two 50-frame API jobs and real Chromium upload/playback, with zero JavaScript errors. Anonymous public pull and actual GPU/CPU-mode video jobs both passed. Native detection results remained unchanged in the container.
The two GPU API jobs took 9.73/9.43 s. Sampled cgroup RAM peaked at 1,025 MiB, whole-device VRAM at 153 MiB; CPU averaged 184%, peaked at 338%, where 100% is one core. Eighteen samples cover the API/worker/FFmpeg phase; peaks can be missed. The image is ~2.20 GB compressed in the registry and reports 6.68 GB locally. These are development observations on one setup. The heuristic attention score has no human attention-accuracy estimate, and the synthetic crossing limitation remains visible. See the linked evaluation report for the full method and failures.
Content type
Image
Digest
sha256:bc953672a…
Size
2 GB
Last updated
23 days ago
docker pull bmaged23/multi-person-attention-tracking:sha-63118083303550468225a2ca9966cfb17ee39b98