Skip to content

Latest commit

 

History

History
 
 

yolov3

README.md

YOLO V3 Object Detection

Sample scripts for training and inferencing of the YOLO V3 object detection model using TensorFlow on DirectML.

This is a fork of https://github.com/zzh8829/yolov3-tf2 at commit 65294d5. The original code is Copyright (c) 2019 Zihao Zhang, and is used here under the terms of the MIT License. See LICENSE for more information.

Setup

First, make sure you have the TensorFlow on DirectML package installed (see GPU accelerated ML training). Then install the pip packages required by YOLO V3:

pip install -r requirements.txt

Next, run the setup script to download all data and pre-trained weights.

python setup.py

Inference

There are two scripts for inference: detect.py (for single images) and detect_video.py (for video frames). You can also stream frames from a webcam.

# Image
python detect.py --image <path>

# Video
python detect_video.py --video <path>

# Webcam
python detect_video.py --video 0

Press the Q key to close the output window and stop the script. Sample images and videos can be found under data.

There are a few additional parameters you may find useful:

  • --output <path> : output the image/video with labels and boxes applied (e.g. --output processed.avi)
  • --headless : don't render output on screen, which is useful for tracing and WSL
  • --trace : emit chrome traces for each image/frame
  • --max_frames <int> : cap on the number of video frames to process (detect_video.py only)

The inference scripts support tracing. For example, to trace first 10 frames of a video without rendering it on screen:

python detect_video.py --video data/grca-trainmix_1280x720.mp4 --trace --max_frames 10 --headless

Training

It takes quite a bit of time, and a good dataset, to properly train YOLO. You can use the train_voc.py helper script to train the model from scratch using the VOC 2012 dataset.

python train_voc.py [--epochs <int>] [--batch_size <int>] [--trace]

Note that YOLO V3 can take up quite a bit of memory, so you may need to reduce the batch size to something smaller than the default (4). Try a --batch_size of 1 or 2 if you run out of memory.