video_editing_poc/docs/local-cv-visual-analysis-gu...

139 lines
3.6 KiB
Markdown

# Local CV Visual Analysis Guide
This service can use a local computer-vision worker for highlight visual analysis.
## Default Behavior
By default the service uses the built-in heuristic provider:
```yaml
video-clipping:
editing:
visual-analysis:
provider: heuristic
```
This keeps the app runnable without Python, GPU drivers, model downloads, or a separate service.
## Enable A Local CV Worker
Run a local worker that accepts JSON over HTTP, then start this Spring service with:
```yaml
video-clipping:
editing:
visual-analysis:
provider: local-cv
endpoint: http://127.0.0.1:8091/v1/analyze-visuals
timeout-ms: 30000
fallback-to-heuristic: true
```
Equivalent environment variables:
```bash
VIDEO_EDITING_VISUAL_ANALYSIS_PROVIDER=local-cv
VIDEO_EDITING_VISUAL_ANALYSIS_ENDPOINT=http://127.0.0.1:8091/v1/analyze-visuals
VIDEO_EDITING_VISUAL_ANALYSIS_TIMEOUT_MS=30000
VIDEO_EDITING_VISUAL_ANALYSIS_FALLBACK_TO_HEURISTIC=true
```
This repository includes an optional starter worker:
```bash
tools/run_local_cv_worker.sh
```
The starter worker installs `tools/local_cv_requirements.txt` into `.venv-local-cv`.
It uses OpenCV for blur, exposure, and face-presence signals. It uses YOLO through `ultralytics` for object detection.
Default model behavior:
- `LOCAL_CV_YOLO_MODEL` defaults to `yolov8n.pt`.
- Ultralytics downloads/caches `yolov8n.pt` on first use if it is not already present.
- Set `LOCAL_CV_YOLO_MODEL=/absolute/path/to/model.pt` to use local weights.
- Set `LOCAL_CV_DISABLE_YOLO=true` to run only OpenCV-based analysis.
Health check:
```bash
curl http://127.0.0.1:8091/health
```
## Worker Request Contract
The Spring service sends:
```json
{
"source": {
"clipId": "porsche-drive",
"sourcePath": "output/highlight-projects/porsche-drive/source/porsche.mp4",
"durationSeconds": 42.0,
"videoCodec": "h264",
"audioCodec": "aac",
"width": 1920,
"height": 1080,
"frameRate": 30.0
},
"thumbnails": [
"output/highlight-projects/porsche-drive/analysis/frames/porsche_0001.jpg"
],
"shotSegments": [
{
"shotId": "shot_0001",
"startSeconds": 0.0,
"endSeconds": 8.0,
"durationSeconds": 8.0,
"representativeTimestampSeconds": 4.0
}
]
}
```
## Worker Response Contract
The worker must return:
```json
{
"clipId": "porsche-drive",
"blurScore": 0.81,
"exposureScore": 0.67,
"motionScore": 0.73,
"compositionScore": 0.79,
"facePresence": "faces_detected",
"objectLabels": [
{
"label": "car",
"confidence": 0.91,
"source": "yolo"
},
{
"label": "luxury sports car",
"confidence": 0.84,
"source": "clip"
}
],
"representativeThumbnails": [
"output/highlight-projects/porsche-drive/analysis/frames/porsche_0001.jpg"
],
"analysisMethod": "local_cv_yolo_clip_mediapipe"
}
```
Scores must be normalized from `0.0` to `1.0`.
## Recommended Local Model Split
- Use `YOLO` for object detection such as `car`, `person`, `food`, `dog`, `bottle`, and scene objects.
- Use `CLIP` or `SigLIP` for semantic labels such as `luxury sports car`, `family birthday`, or `restaurant dish`.
- Use `MediaPipe` or OpenCV Haar cascades for face presence when lightweight local face detection is enough.
- Use OpenCV/Laplacian variance and brightness histograms for blur and exposure scores.
## Runtime Behavior
When `fallback-to-heuristic=true`, the service logs `event=local_cv_visual_analysis_fallback` and writes heuristic visual analysis if the CV worker is down or returns an error.
When `fallback-to-heuristic=false`, local CV failures fail the source analysis and the source video is rejected by the scheduler.