Egocentric Video Processing & Quality Control Platform
Separating useful work footage from noise in hours-long first-person video — with pose-level QC.
The problem
Long egocentric recordings of people working contain large stretches of irrelevant footage, and the human pose data extracted from them is only useful if it is complete and temporally consistent. Reviewing either by hand doesn't scale.
The approach
A pipeline that samples long videos, uses a vision-language model to label segments KEEP / REMOVE / REVIEW, cuts clean clips with FFmpeg, then estimates and tracks human pose and runs skeleton quality checks — with a review dashboard for everything the system isn't sure about.
Architecture
System architecture
Stage 1 / 11
Upload
Long-duration videos are uploaded and registered with their metadata.
Target architecture of a project in active development — not every stage is complete.
Key features
- Long-video upload and preprocessing
- VLM-assisted KEEP / REMOVE / REVIEW decisions
- Automatic clip filtering with FFmpeg
- Pose estimation and person tracking
- Skeletal keypoint and temporal consistency checks
- Human review dashboard
- Export of clean video and structured metadata