:unicorn: YOLOv3 Implementation in TensorFlow 1.1x + Keras :unicorn:

How it Looks Like

Watch the demo

Quick Start

On a PC / Mac

Create conda environment depending on whether you have a supported GPU or not:

conda env create -f environment-[c|g]pu.yml
source activate yolov3-[c|g]pu

On a Raspi 3

Install OpenCV 3 with the following instructions.

Then:

pip install tensorflow scikit-learn

Download YOLO Weights

Download weights into the cfg directory:

cd cfg
wget https://pjreddie.com/media/files/yolov3.weights

Demo on Single Image:

python single_image.py

The output is stored on out.png in the root folder.

Demo on Web Cam:

To see it live on your Web Cam:

python webcam.py

Progress

• YOLO configuration parser
• Build YOLO model
• Check architecture against a well-known implementation
• Load YOLO pre-trained weights
• Handle YOLO layer (Detection Layer)
• Non-Maximal Suppression
• Colorful boxes with labels and scores
• Test out on a Web Cam
• Check dependencies
• Dependencies for CPU and GPU
• Instructions for running the project
• Use original scale of input image
• YOLO head in a function
• Figure out Eager Execution + Loading Weights
• Support Tiny YOLOv3
• Allow passing in options to use Tiny YOLOv3
• Investigate Quantization / Smaller Weights
• Try this out on a Raspi3
• Tensorflow.js (¯\(ツ)/¯)

Credits

• Series: YOLO Object Detector in PyTorch by Ayoosh Kathuria
• Implementing YOLO v3 in Tensorflow (TF-Slim) by Paweł Kapica
• xiaochus/YOLOv3
• qqwweee/keras-yolo3
• kevinwuhoo/randomcolor