Camera-based object detection, recorded demos

A running archive of detection work: point a camera at an object, the model identifies it, the overlay shows what it found and how sure it is. Each clip below is a recording of a working build.

Detection model by Alejandro Lopez Cordero Site built by Erick Ramirez View the archive
3 steps YOLO26s Single class: gun Updated Aug 2026

Step 1 / Input

The frame is resized and converted into numbers so the neural network can read it.

YOLO-GunRecognizer

Deep Learning · Computer Vision AI

A single-class YOLO model (class: “gun”) built for live-monitoring and real-surveillance frames, using the YOLO26s architecture. Improved 3.8x through downsampling, hard-negative integration, and data-leakage inspection, then evaluated against the UGR and SOHAS benchmarks.

  • Eval UGR benchmark: P 0.922 R 0.787 mAP50 0.850 mAP50-95 0.637 | R@0.2 0.824 P@0.865 real-CCTV-recall improved from 0.149 (baseline) to 0.568
  • Eval SOHAS: P 0.710 R 0.800 mAP50 0.787 mAP50-95 0.654 | R@0.2 0.952 P@0.2 0.400
  • Deep Learning
  • YOLO
  • Pytorch
  • Roboflow
  • SLURM

Model by Alejandro Lopez Cordero

View repository

Step 2 / Processing

A single frame with a detection box labelled gun 0.90 drawn around a rifle.

Convolutional layers scan the image and predict bounding boxes with a confidence score for each.

Inference

ONNX Runtime · In-browser

Every frame is letterboxed to 960×960 with grey padding, the same preprocessing used at training time, so aspect ratio holds and boxes come back unskewed. The export is end-to-end: NMS sits inside the graph, so the model emits a fixed [1, 300, 6] tensor instead of raw proposals.

  • M1 Max, WebGPU: median inference 40 ms, spread 39–42 ms, first pass 51 ms including shader compile
  • End to end: 19.2 fps on a 1920×1080 source; without WebGPU it falls back to single-threaded WASM
  • ONNX Runtime
  • WebGPU
  • WASM
  • 960×960
Run it live

Step 3 / Output

The box with the highest confidence is kept and drawn around the detected firearm.

Detections

Thresholding · Overlay

The page thresholds that tensor and draws whatever survives. One class only — index 0 is gun, so there is no label table to look up. Confidence is adjustable while it runs, which makes the precision-recall trade visible rather than theoretical.

  • Model output: fixed [1, 300, 6] tensor per frame, already NMS-filtered in-graph
  • Weights: 38 MB, fetched once and held in the Cache API, so a revisit does no network I/O
  • Single class
  • In-graph NMS
  • Confidence
Try it yourself