WSDDN PyTorch

Implementation of Weakly Supervised Deep Detection Networks using the latest version of PyTorch.

Bilen, H., & Vedaldi, A. (2016). Weakly supervised deep detection networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 2846-2854).

Implementation Differences

• Adam optimizer (instead of SGD)
• Spatial regulariser isn't added

Experiments

• VGG16 based model is closest to EB + Box Sc. case with model L, which reported 30.4 mAP in the paper
• AlexNet based model is closest to EB + Box Sc. case with model S, which reported 33.4 mAP in the paper
• Results when VGG16 is used as base model

aero	bike	bird	boat	bottle	bus	car	cat	chair	cow	table	dog	horse	mbike	person	plant	sheep	sofa	train	tv	mean
41.4	46.3	22.7	24.5	13.6	57.7	49.9	31.1	7.5	31.1	24.3	25.9	38.7	53.5	7.2	13.9	31.1	38.6	48.3	39.0	32.3

• Results when AlexNet is used as base model

aero	bike	bird	boat	bottle	bus	car	cat	chair	cow	table	dog	horse	mbike	person	plant	sheep	sofa	train	tv	mean
38.1	41.5	27.1	18.6	10.3	48.8	47.6	36.8	1.6	25.9	28.5	30.4	39.7	46.8	15.1	12.4	28.3	32.4	44.2	44.8	30.9

Requirements

• Docker (19.03.2)
• nvidia-container-toolkit (https://github.com/NVIDIA/nvidia-docker)

Build Steps

git clone git@github.com:adursun/wsddn.pytorch.git
cd wsddn.pytorch
./prepare.sh
docker run --rm --gpus all --ipc=host -v `pwd`:/ws -it wsddn.pytorch /bin/bash

Training Steps

# for VGG based model
python src/train.py --base_net vgg

Evaluation Steps

# for VGG based model
# run `wget "https://www.dropbox.com/s/xyi4hgms6y3ldmj/vgg_epoch_20.pt?dl=1" -P states/` to use pretrained weights
python src/evaluate.py --base_net vgg --state_path states/vgg_epoch_20.pt