Edge-Case Miner
Find the rare edge-case clips AV and robotics teams need for training: describe a scenario in plain English, search every camera in a VAST video archive, let NVIDIA Cosmos3-Reason watch each candidate, and export a verified, versioned dataset.
How it works
Describe the edge case in plain English, or pick one of 10 taxonomy scenarios.Expand. Qwen 3.8 on W&B Inference rewrites it as 4 caption-style search queries.Search. The scenario and its queries run concurrently against VSS hybrid search (Cosmos-Embed1 text + visual vectors in VastDB) across every camera. Hits are merged (best similarity, which queries found them) and shown with their YOLO11 object counts.Verify. Cosmos3-Reason watches each candidate: the app pulls the 5 s segment, shrinks it to 480p / 8 fps H.264 and asks for a strict JSON verdict {match, confidence, why}. If Cosmos is unavailable, DeepSeek V4 on W&B Inference judges the ingest caption instead (labelled "caption"). Verdicts are cached.Grow from good hits. "More like this" searches with a clip's caption; before/after steps to the neighbouring segments of the same video.Coverage. A scenario × location grid. Cells with no verified match are GAPs (red): the edge cases to go collect or simulate.Export. The verified set becomes a manifest (S3 URI, timestamps, camera, location, label, verdict, reason) and a versioned W&B Artifact. Every step is traced in W&B Weave.
