Efficient Decoder-free Object Detection with Transformers

ECCV2022: Paper

This code is based on mmdetection

Results

Backbone	box mAP	GFLOPs
$DFFT_{NANO}$	42.8	42
$DFFT_{TINY}$	43.5	57
$DFFT_{SMALL}$	44.5	62
$DFFT_{MEDIUM}$	45.7	67
$DFFT_{LARGE}$	46.0	101

Models

ImageNet Pretrained

MS COCO Detection

Runs

python3 setup.py develop && ./tools/dist_train.sh configs/dfft/dfft.py [num_gpus] \
--cfg-options model.pretrained=pretrained.pth