Face detection and redaction expert: paid 1-week trial to fix face blurring in our video pipeline

FACE DETECTION AND REDACTION EXPERT: PAID 1-WEEK TRIAL TO FIX FACE BLURRING IN OUR VIDEO PIPELINE We record first-person video with a head-mounted GoPro (4K, 30 fps). Every face must be blurred when it comes out of our post-processing pipeline, and only the face: anything blurred beyond it is footage we cannot use. Our Python pipeline handles ingest, review and packaging. The face detector and the blur that follows it are the part that does not work. That is the job. THE PROBLEM The pipeline runs OpenCV YuNet at 0.9. It blurs the faces it finds well and misses reflections. On a measured batch (8 clips, 107 minutes, 8 spans where the wearer's face shows in a reflection) YuNet caught 2 of 8. RetinaFace caught 6 of 8 with 1 false alarm on 300 frames that contain no face, at 1.55 seconds per frame on a 4-core CPU. Nothing we tried found faces behind glass shower doors. The misses are covered by hand today: a reviewer lists the span and the pipeline blurs the whole frame for it. We need the detector to find those faces so that only the face is blurred and the whole-frame fallback goes away. What to expect in the footage: the wearer reflected in bathroom mirrors, fridge and microwave doors and shower glass; photos of people on a fridge or a phone lock screen; occasionally another person, including children. Whole episodes may contain no face at all, so false alarms cost as much as misses. Domestic interiors are full of things that look like faces to a detector: packaging, toys, cartoon prints, framed photos, TV screens. What counts as a face: a real person, including reflections, photos and people on screens. Cartoons, toys and artwork are not faces, and blurring them counts as a false alarm. THE TARGET - Recall at least 99%: every real face blurred in every frame it appears. - Precision at least 95%: at least 95 in 100 blurred boxes are real faces. - Once both are met, the lowest total cost per video-hour, GPU included. THE TRIAL (FIXED PRICE, $500, 1 WEEK) We will pick 3 candidates from the proposals and give each the same task. The best result wins the ongoing work. What you get at contract start: - Access to our pipeline's code repository. - 15 real clips, 4K HEVC. - Our answer key for all 15: a face_timestamps.csv listing when each face appears. What you deliver within 7 days: 1. Test models and settings to find the best solution, and run it on all 15 clips. 2. The blurred videos from your best solution, for all 15 clips. 3. Your measured recall and precision on the 15 clips, the cutoff you recommend and why, the hardware you used, and the cost per video-hour. 4. A pull request that updates our pipeline to use your model, with instructions to run it and the hardware it needs. 5. A list of every pretrained weight you used, with its licence. Weights must allow commercial use. The model and code are ours. How we score: we also run your pull request on clips you haven't seen and score the result. Every face moment must be blurred across its whole span, at the face only, with no blur outside face moments. The answer key is for checking your work, not an input: your solution must find faces without it. AFTER THE TRIAL (HOURLY, $30 TO $100, AT OUR OPTION) About 10 hours a week through January 2027 for the person we keep: improving the detector as new footage arrives, and detectors for screens, documents and addresses. WHAT WE NEED FROM YOU - You have trained or fine-tuned a face detector (or another detector) on your own data and can link to it with the number it moved. - You have shipped face anonymisation in video and can say what makes it irreversible and how you kept the surrounding footage intact. - You can start within a week, and have at least 3 hours a day of overlap with US Central Time. TO APPLY, ANSWER THE FOUR SCREENING QUESTIONS. We read links, numbers and Upwork work history, not cover letters; a proposal with nothing we can check will be skipped.