Problem
Avito added video to listings and had to moderate it. Manual video moderation is expensive, slow, hard to scale with the flow and error-prone. Automating it started from poor data: video was new, violations are a small share, the labels were noisy and had no violation reasons, and videos with violations were sometimes deleted from storage.
Example from the article: an apartment listing with a video about May 1 and labor, caught as "video not about the item".
Solution in brief
Split a violation into a domain (frames, audio, metadata, link to the listing) plus a reason, and cover each domain with a simple, reusable detector: CLIP frame embeddings with a binary head per reason, OCR and ASR with regexes, logo and face matching. Measure online by sending a sample of "no violation" videos to manual review and scaling the result up.
Questions this case answers
- How to launch automatic moderation of uploaded content with many violation reasons when labeled violations are scarce and noisy
- How to estimate precision, recall and error rate of an automatic filter when manual review only sees a sample of the items it let through
- How to launch automatic moderation of user-uploaded video when labeled violations are scarce and noisy
- How to cover a new violation reason quickly without training data
Tools and methods
- CLIP — frame embeddings, zero-shot detection and text-to-video matching
- Joint Net — merges frame embeddings into one video embedding (Avg, Sum, Attention, CNN, RNN)
- OCR and Whisper ASR with regular expressions — violations in on-screen text and speech
- Detectors and matchers for logos and faces — compare embeddings with a database
- Gradient boosting on metadata — filters out very low-quality video by bitrate and FPS
- Sampler for manual review — estimates false negatives and true negatives
Google or a one-time email code. No password.