Tensorflow YOLOv3


YOLOv3 and YOLOv3-Tiny Implementation for Real-Time Object Detection in Tensorflow

This is my first project in Computer Vision. I wanted to compare both YOLOv3 and YOLOv3-Tiny performance.

Later I will do a Transfer Learning for a future project.


Run make install

Then download the official weights.

For YOLOv3:

curl https://pjreddie.com/media/files/yolov3.weights > ./weights/yolov3.weights

For YOLOv3-Tiny:

curl https://pjreddie.com/media/files/yolov3-tiny.weights > ./weights/yolov3-tiny.weights


python3 convert_weights.py [-h] [--tiny]

python3 detect.py [-h] [--tiny] {video,image} iou confidence path


First we convert weights for YOLOv3 model:

python3 convert_weights.py

Instructions for converting weights are in ./convert_weights.py.

Then we can run detection for either an images:

python3 detect.py image 0.5 0.5 ./data/images/person.jpg

Or for a video:

python3 detect.py video 0.5 0.5 data/videos/champs-elysees.mp4

YOLOv3-Tiny vs. YOLOv3


The YOLOv3-Tiny boxes that use the smaller anchors have the right midpoint but sometimes the width / height seems off.


A Tensorflow implementation

YOLOv3 Paper

Video Source

Project done in 2019