This package is a from-scratch implementation of YOLOv3 in PyTorch capable of running in real time on webcam streams as well as on image files and video files. It parses the original Darknet configuration and weights files to build the network and has been tested with the yolov3, yolov3-tiny, and yolov3-spp models.

Link to webcam demo on YouTube

A more detailed treatment of YOLOv3 and the code in this repo can be found in the blog posts at nrsyed.com

Requirements

• Python ≥ 3.6

Installation

Clone the repo, install the package, and download weight files to the models directory:

git clone https://github.com/nrsyed/pytorch-yolov3.git
cd pytorch-yolov3
pip install .
./get_weights.sh

Examples

Refer to the Usage section for a complete list of options, or run yolov3 --help. See the doc/img directory for example image and webcam video output.

# Run yolov3 on a webcam stream in real time (using the GPU and CUDA)
yolov3 -C -c models/yolov3.cfg -w models/yolov3.weights -n models/coco.names

# Run yolov3-tiny on a webcam stream from webcam 2 (using the CPU instead
# of GPU) and display the FPS in real time
yolov3 -C 2 -c models/yolov3-tiny.cfg -w models/yolov3-tiny.weights \
  -n models/coco.names -d cpu --show-fps

# Run yolov3-spp on a directory of images using the computer's 2nd GPU and
# only display objects with a probability score >0.15
yolov3 -I sample_dataset/images -c models/yolov3-spp.cfg \
  -w models/yolov3-spp.weights -n models/coco.names -d cuda:1 -p 0.15

# Run yolov3 on a video file and write the resulting video (with bboxes) to
# a new file.
yolov3 -V /path/to/video.mpeg -c models/yolov3.cfg -w models/yolov3.weights \
  -n models/coco.names -o /path/to/output.mp4

Usage

Required arguments

One of the following input source arguments is required:

• -C/--cam [cam_id]: Webcam device ID or path to video stream (e.g., RTSP URL); defaults to 0 if none provided.
• -I/--image <path>: Path to an image file or directory of images.
• -V/--video <path>: Path to a video file.

The following arguments are always required:

• -c/--config <path>: Path to Darknet model .cfg file.
• -w/--weights <path>: Path to Darknet model .weights file.

Optional arguments

• -d/--device <device>: Device on which to load the model (e.g., cpu, cuda, cuda:1)
• -h/--help: Display help message and exit.
• -i/--iou-thresh <iou>: Non-maximum suppression IOU threshold.
• -n/--class-names <path>: Path to text file of class names containing one class name per line. If omitted, the class indices are displayed next to the bounding boxes in the resulting image instead of the class names.
• -p/--prob-thresh <prob>: Detection probability threshold; predictions with a computed probability below this threshold are ignored.
• -o/--output <path>: Path for output video file (use with --cam or --video input source options). Only .mp4 filetype supported. For --cam, the output video framerate (FPS) is equal to the average framerate over the duration of the recording. For --video, the output video framerate is equal to the input file framerate.
• --show-fps: Display frames processed per second (--cam input only).

Acknowledgments

The model config files in the models directory were obtained from https://github.com/pjreddie/darknet/tree/master/cfg. The list of COCO class names in the models directory was obtained from https://github.com/pjreddie/darknet/tree/master/data and modified to match the COCO annotation category names. Weights for yolov3, yolov3-tiny, and yolov3-spp can be found on Joseph Redmon's website at https://pjreddie.com/darknet/yolo/. Sample images were taken from the original COCO dataset (http://cocodataset.org); the functions in the yolov3.devtools module were used to select images containing at least 6 classes of objects belonging to at least 6 distinct supercategories and no more than 10 objects (from the 2017 validation images dataset).