This is the official pytorch implementation for paper: BiDet: An Efficient Binarized Object Detector, which is accepted by CVPR2020. The code contains training and testing stages on two binarized object detectors, SSD300 and Faster R-CNN on PASCAL VOC and Microsoft COCO 2014.

• python 3.5+
• pytorch 0.4.0+
• other packages include numpy, cv2, matplotlib, pillow, cython, cffi, msgpack, easydict, pyyaml

We conduct experiments on PASCAL VOC and Microsoft COCO 2014 datasets.

We train our model on the VOC 0712 trainval sets and test it on the VOC 07 test set. For downloading, just run:

sh data/scripts/VOC2007.sh # <directory>
sh data/scripts/VOC2012.sh # <directory>

Please specify a path to download your data in, or the default path is ~/data/.

We train our model on the COCO 2014 trainval35k subset and evaluate it on minival5k. For downloading, just run:

sh data/scripts/COCO2014.sh

Also, you can specify a path to save the data.

After downloading both datasets, please modify file faster_rcnn/lib/datasets/factory.py line 24 and file faster_rcnn/lib/datasets/coco.py line 36 by replacing path/to/dataset with your voc and coco dataset path respectively.

The backbones for our BiDet-SSD300 and BiDet-Faster R-CNN are VGG16 and Resnet-18. We pretrain them on the ImageNet dataset. You can download the pretrained weights on: VGG16 and ResNet18. After downloading them from Google Drive, please put them in ssd/pretrain and faster_rcnn/pretrain respectively.

Assume you've finished all steps above, you can start using the code easily.

For training SSD, just run:

$ python ssd/train_bidet_ssd.py --dataset='VOC/COCO' --data_root='path/to/dataset' --basenet='path/to/pretrain_backbone'

For testing on VOC, just run:

$ python ssd/eval_voc.py --weight_path='path/to/weight' --voc_root='path/to/voc'

For testing on COCO, just run:

$ python ssd/eval_coco.py --weight_path='path/to/weight' --coco_root='path/to/coco'

First you need to compile the cuda implementation for RoIPooling, RoIAlign and NMS. Just do:

cd faster_rcnn/lib
python setup.py build develop

For training Faster R-CNN, just run:

$ python faster_rcnn/trainval_net.py --dataset='voc/coco' --data_root='path/to/dataset' --basenet='path/to/pretrain_backbone'

For testing, run:

$ python test_net.py --dataset='voc/coco' --checkpoint='path/to/weight'