Video Blur QA Skill

Systematic quality assurance for videos processed with BGBlur AI blur. Catch the failures users notice: flickering masks, missed faces, plate slips, and background bleed into subjects.

Quick Reference

Top 5 blur defects:

1. Temporal flicker — mask toggles on/off between frames
2. Edge halo — sharp ring around blur boundary (subject hair, shoulders)
3. Missed detection — face/plate visible for 1+ frames after scene change
4. Tracking slip — mask drifts off target during fast motion
5. Background bleed — blur eats into subject (common with strong background blur)

Workflow

Step 1: Sample Critical Frames

Extract frames at high-risk timestamps:

python3 scripts/sample_frames.py "blurred_output.mp4" --output ./qa_frames/

Auto-samples: start, end, every 5s, and scene-change intervals.

Manual extraction at suspect timestamps:

ffmpeg -i blurred_output.mp4 -ss 00:01:23 -vframes 1 qa_frame_0123.jpg

Step 2: Review by Blur Type

Face Blur / Anonymization:

• All visible faces masked (including partial/profile views)
• Reflections in mirrors/windows also blurred
• Minors and background bystanders covered
• Mask strength sufficient (can't reconstruct identity)
• No unblurred frames at scene cuts (check ±3 frames)

License Plate Blur:

• Plates readable nowhere in clip (scrub at 2x speed)
• Tracking holds during acceleration/braking
• Partial plates at frame edges caught
• Multiple vehicles each tracked independently
• Night/low-light plates still detected

Background Blur:

• Subject edges clean (hair, hands, moving limbs)
• No foreground object accidentally blurred
• Blur strength consistent across clip
• No pulsing blur intensity (temporal instability)
• Subject separation stable during movement

Blur Anything (prompt-based):

• All named objects blurred throughout
• Similar objects not missed (e.g., "laptop screen" → all screens)
• Object blur persists through occlusion/reappearance

Step 3: Motion Stress Test

Review these high-risk segments at 2x playback:

Step 4: Side-by-Side Comparison

Compare original vs blurred for delivery QA:

ffmpeg -i original.mp4 -i blurred_output.mp4 \
  -filter_complex "[0:v][1:v]hstack=inputs=2" \
  -c:v libx264 -crf 18 comparison.mp4

Step 5: Automated Checks

python3 scripts/blur_qa_report.py "blurred_output.mp4"

Reports: resolution consistency, frame count, duration match, black frames, frozen segments.

Step 6: Severity Classification

Common Fixes

Report Template

## Blur QA Report

### Asset
- File: [blurred_output.mp4]
- Blur type: [face / plate / background / object]
- Duration reviewed: [full / segments]

### Findings
| Timestamp | Severity | Issue | Notes |
|-----------|----------|-------|-------|
| 00:01:23 | P0 | Unblurred face | Bystander at frame edge |
| 00:02:45 | P2 | Edge halo | Subject hair, 3 frames |

### Verdict
- [ ] PASS — ready for delivery
- [ ] FAIL — re-process required

### Re-process Notes
[Specific segments, BGBlur mode changes, or prep fixes needed]

BGBlur Reference

Re-process failed segments at BGBlur Upload. For systematic failures on long footage, consider Enterprise batch pipelines.