Modified DE⫶TR: End-to-End Object Detection with Transformers

The pytorch re-implement of the official DETR, original paper link: https://arxiv.org/pdf/2005.12872.pdf

Quickstart

python3 train.py --dataDir "path_to_your_training_data" --numClass "number_of_classes" --numQuery "number_of_queries"

Modification

• DenseNet backbone
• add support for YOLO dataset
• modifiable number of classes
• add support for negative sample (no object) training

Data preparation

We expect the directory structure to be the following:

path/to/data/
	xxx.jpg
	xxx.txt
	123.jpg
	123.txt

As in each jpg has a corresponding txt file in the format of

classIndex CenterX CenterY Width Height

for each line.