# Object Detector

> Free, in-browser object detection for photos and a live webcam. Boxes and counts for the 80 COCO object classes (people, vehicles, animals, bags, food, furniture, electronics). Runs on the visitor's device; nothing is uploaded, no sign-in, no credits.

URL: https://object-detector.skillsafe.ai/

## What it does
- Photos: drop, paste or choose one or many images; each gets boxes, per-class counts and a per-photo table.
- Webcam: live detection from the front or back camera, with frames per second and inference time.
- Count over time: for the webcam, the largest count of each object in any frame of each second, charted for the last two minutes (table view and CSV of up to ten minutes).
- Controls: Fast (YOLOX-nano, 416 px, 3.7 MB) or Accurate (YOLOX-s, 640 px, 36 MB) model; confidence threshold; overlap (NMS IoU) threshold; filter by class, by group or by search; labels and scores on or off; click a class to highlight its boxes.
- Exports: PNG snapshot with boxes, CSV of every box (image, object, confidence, x1, y1, x2, y2, width, height, in the original photo's pixels), per-photo counts CSV, COCO-format JSON (images, annotations with [x, y, w, h] boxes and scores, categories with official COCO ids), counts copied as text.
- Two example photos show precomputed results instantly; "Run on this device" re-runs them live.

## How it works
YOLOX (Megvii, Apache-2.0) in onnxruntime-web 1.27, WebGPU when available, otherwise WebAssembly on the CPU. Model files come from SkillSafe's model host, are verified against their SHA-256 and kept in the browser's Cache Storage. Boxes are decoded from the anchor grid (strides 8/16/32), scored as objectness x class score, and suppressed per class with greedy non-maximum suppression.

## Limits
Only the 80 COCO classes. Small or heavily occluded objects are missed more often by the Fast model. Counts are of detections, not of tracked identities.

## Credits
YOLOX: Copyright (c) 2021-2022 Megvii Inc., Apache License 2.0. Example photos: Wikimedia Commons, CC0 ("Moscow bicycle parade 2025-05-17 (136)", "Moscow, Hummer limo, Zhongtong bus July 2026 02"). Runtime: onnxruntime-web, MIT.