A Keras implementation of YOLOv3 (Tensorflow backend) forked from qqwweee. Main contribution over the previous version is that this is able to do batch inference now. This comes with a caveat that the input image size needs to be pre-defined now.

• Download from qqwwee yolov3.h5

python convert.py yolov3.cfg yolov3.weights model_data/yolo.h5

• yolo.py is intended for use as an object, for example from yolo import YOLO, then instantiate the YOLO class as an object.

• Take note of the default parameter by looking into yolo.py. One important parameter often overlooked is the input image size.

• There are many methods in the class due to different projects needing it and legacy reasons, but the main method to use is detect_get_box_in where you give a list of ndarray-like images and can specify the format you want the BBs back in.

Take a look at example_video.py on how I will use it on a video

This implementation is special/weird in the sense that while the inference is in Keras, but the preprocessing is done with tensorflow aka computation graphs.