ASSD-Pytorch

Please cite the article in your publications if it helps your research (Arxiv Link, ELSEVIER Link):

@article{YI2019102827,
    title = "ASSD: Attentive single shot multibox detector",
    journal = "Computer Vision and Image Understanding",
    pages = "102827",
    year = "2019",
    issn = "1077-3142",
    doi = "https://doi.org/10.1016/j.cviu.2019.102827",
    url = "http://www.sciencedirect.com/science/article/pii/S1077314219301328",
    author = "Jingru Yi and Pengxiang Wu and Dimitris N. Metaxas",
}

ASSD learns to highlight useful regions on the feature maps while suppressing the irrelevant information, thereby providing reliable guidance for object detection.

System	VOC2007 test	VOC2012 test	FPS (TitanX)	#Boxes	Input resolution
SSD300 (VGG16)	77.2	75.8	46	8732	300 x 300
SSD512 (VGG16)	79.8	78.5	19	24564	512 x 512
ASSD300 (VGG16)	80.0	77.5	-	8732	300 x 300
ASSD321 (ResNet101)	79.5	76.4	27.5	10325	321 x 321
ASSD512 (VGG16)	81.6	80.0	-	24564	512 x 512
ASSD513 (ResNet101)	83.0	81.3	16	25844	513 x 513

Dependencies

Library: OpenCV-Python, PyTorch>0.4.0, Ubuntu 14.04

Dataset

PascalVOC

# Download the data.
cd $HOME/data
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
# Extract the data.
tar -xvf VOCtrainval_11-May-2012.tar
tar -xvf VOCtrainval_06-Nov-2007.tar
tar -xvf VOCtest_06-Nov-2007.tar

MSCOCO 2017 (download link)

  #step1: download the following data and annotation
  2017 Train images [118K/18GB]
  2017 Val images [5K/1GB]
  2017 Test images [41K/6GB]
  2017 Train/Val annotations [241MB]
  #step2: arrange the data to the following structure
  COCO
  ---train
  ---test
  ---val
  ---annotations

Train/Test/Evaluation

1. Change the mode in main.py
2. Change parameters such as root (data directory) in config.py
3. python main.py