kevinz-code / yolov1

A Pytorch Implementation of YOLOv1

YOLOv1

For better performance, this not the same as the original paper.

I achieved 0.684 mAP on VOC07test, 76fps on RTX2080Ti

For better training speed, I changed the backbone from vgg to resnet50. And add a few 1x1 and 3x3 conv to fine-tune the resnet. For better detection for small objects, I change the 7x7 feature maps to 14x14 feature maps and drop the fully connected which has been implemented in the original paper.

Trained on VOC2007+VOC2012

model	map on VOC07test	FPS
YOLO Resnet50	68.4%	76
YOLO original	63.4%	45

Prerequisites

• pytorch 1.2.0
• cuda 10.0.1
• pillow 6.2.1
• numpy

Quick Start

Download the file

git clone https://github.com/Kevinz-code/YOLOv1.git
python demo.py

And You will find the demo results pictures in

./demo/demo_results

Training

1. Download VOC2012train and VOC2007train dataset
2. Download VOC2007test dataset
3. Put them in the dir

../Image/

To train from scratch, run

python main.py -s 0 

To get parameters help, run

python main.py -h

This will automatically start train on VOC07+12, and test on VOC07 every epochs.

Details

Some parameters setting are very Important. And I spent a long time trying to find these best parameters. For convenience, I list them below.

For Training

learning rate	3e-3, 1e-3
weight_decay	0.0005
miniBatch	16
epoch	30
momentum	0.9

For Testing

confidence_thresh	0.3
nms_thresh	0.26