YOLO without the framework bloat

What it does

This is a single-file Python script (yolo_opencv.py) that loads a pre-trained YOLO v3 model and runs object detection on an image using OpenCV’s dnn module. You supply four files: an image, the YOLO config, the weights, and a class list. It draws bounding boxes and labels on the output.

The interesting bit

OpenCV’s DNN module can run inference directly on models from Caffe, Torch, and TensorFlow — including DarkNet/YOLO since a relatively recent addition. That means you can do GPU-accelerated-ish object detection without installing a full deep learning framework stack. The README notes SSD and Faster R-CNN examples are “coming soon,” but only YOLO is implemented here.

Key highlights

• Single dependency stack: just numpy and opencv-python
• Runs on pre-trained YOLO v3 weights downloaded directly from pjreddie.com
• Command-line interface with explicit paths for image, config, weights, and classes
• Includes a sample output image showing detected objects with bounding boxes
• Author’s follow-up library cvlib wraps this in a one-liner (detect_common_objects())

Caveats

• Python 2.x compatibility is explicitly not tested
• SSD and Faster R-CNN examples mentioned but not yet implemented
• No video/webcam support shown in the README; image files only

Verdict

Worth a look if you need a quick, dependency-light YOLO demo or want to understand how OpenCV’s DNN module handles model loading. Skip it if you need production pipelines, real-time video processing, or training capabilities — this is strictly inference glue.