DETR: An End-to-end Object DEtector using TRansformer

About the code

Paddle training code and pretrained models for DETR (DEtection TRansformer). This code is modified from original project to make it compatible with Paddle.

Original project: DETR(pytorch)

Paper: End-to-End Object Detection with Transformers

Aistudio Project

If this work is useful to you, please cite:

@INPROCEEDINGS{DETR,
title={End-to-End Object Detection with Transformers},
author={Nicolas Carion and Francisco Massa and Gabriel Synnaeve and Nicolas Usunier and Alexander Kirillov and Sergey Zagoruyko},
year={2020},
booktitle={ECCV}
}

Contributions:

• This project provided DETR training code in Paddle.
• This project provided pretrained models converted from Pytorch version.
• This project analysis and visualize the model outputs.

Model Zoo

	name	backbone	schedule	inf_time	box AP	url	size	paddle model
0	DETR	R50	500	0.036	42.0	model | logs	159Mb	detr.pdparams
2	DETR	R101	500	0.050	43.5	model | logs	232Mb	detr.pdparams

Train - Object detection

Warning: this code need to optimize, model is very memory-comsuming!!!

Train COCO 2017

1. unzip dataset, run cd /home/aistudio/data/data7122/ && unzip train2017.zip && unzip val2017.zip && unzip annotations_trainval2017.zip.

2. run cd ../../detr && python ./main.py --dataset_file coco

Train Visual Genome

1. unzip dataset, cd /home/aistudio/data/data57396 && zip -s 0 full_images.zip --out image.zip && unzip image.zip && unzip v1.0.zip
2. run cd ../../detr && python ./main.py --dataset_file vg

Train Visual Relationship Detection

! cd /home/aistudio/data/data57355/ && unzip json_dataset.zip
! clear
! cd /home/aistudio/data/data57355/ && unzip sg_dataset.zip
! clear
! cd /home/aistudio/data/data57355/sg_dataset/sg_test_images/ && mv ./4392556686_44d71ff5a0_o.* ./4392556686_44d71ff5a0_o.jpg

2. cd ../../detr && python ./main.py --dataset_file vrd