A PyTorch implementation of a YOLO v1 Object Detector

Implementation of YOLO v1 object detector in PyTorch. Full tutorial can be found here in korean.

Tested under Python 3.6, PyTorch 0.4.1 on Ubuntu 16.04, Windows10.

Requirements

See requirements for details.

NOTICE: different versions of PyTorch package have different memory usages.

How to use

Training on PASCAL VOC (20 classes)

main.py --mode train -data_path where/your/dataset/is --class_path ./names/VOC.names --num_class 20 --use_augmentation True --use_visdom True

Test on PASCAL VOC (20 classes)

main.py  --mode test --data_path where/your/dataset/is --class_path ./names/VOC.names --num_class 20 --checkpoint_path your_checkpoint.pth.tar

pre-built weights file

python3 utilities/download_checkpoint.py

pre-build weights donwload

Supported Datasets

Only Pascal VOC datasets are supported for now.

Configuration Options

argument	type	description	default
--mode	str	train or test	train
--dataset	str	only support voc now	voc
--data_path	str	data path
--class_path	str	filenames text file path
--input_height	int	input height	448
--input_width	int	input width	448
--batch_size	int	batch size	16
--num_epochs	int	# of epochs	16000
--learning_rate	float	initial learning rate	1e-3
--dropout	float	dropout probability	0.5
--num_gpus	int	# of GPUs for training	1
--checkpoint_path	str	checkpoint path	./
--use_augmentation	bool	image Augmentation	True
--use_visdom	bool	visdom	False
--use_wandb	bool	wandb	False
--use_summary	bool	descripte Model summary	True
--use_gtcheck	bool	gt check flag	False
--use_githash	bool	use githash	False
--num_class	int	number of classes	5

Train Log

Results

Authorship

This project is equally contributed by Chanhee Jeong, Donghyeon Hwang, and Jaewon Lee.

Copyright

See LICENSE for details.

REFERENCES

[1] Redmon, Joseph, et al. "You only look once: Unified, real-time object detection." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. (https://arxiv.org/abs/1506.02640)