BadBob Playground · Original MIT reference
Local detection in your browser
Load a compatible ONNX detector and try an image first. Model files, images and camera frames stay in your browser. Runtime assets come from this app's origin.
1. Load your model
One RGB float32 input [1,3,640,640]. One float32 output [1,4+C,N] with pixel-space center XYWH and class probabilities. No objectness or embedded NMS. Match the labels to the model's exact class order.
2. Choose your input
Only Start camera requests video permission. No audio, recording or upload. Camera needs HTTPS or localhost and a browser with OffscreenCanvas, ImageBitmap and WebAssembly SIMD.
Model limits and cleanup
Preprocessing preserves aspect ratio with rounded letterbox dimensions, padding 114 and division by 255. Confidence ≥0.25; class-aware IoU threshold 0.45. Best 3000 candidates, at most 300 detections. Maximum 1024 classes and 100000 candidates. Label count is checked; only you can verify semantic label order.
Stop closes camera tracks immediately, rejects stale results and requests session release. A running WebAssembly call cannot be interrupted cooperatively; after five seconds the app terminates its worker. Replacing a model closes the previous worker before starting another. Reload your model after stopping.
Use only trusted models. This strict adapter does not support arbitrary ONNX models, Core ML files, logits, objectness, segmentation or embedded NMS.