Segmentation-Aggregation Framework for Weakly Supervised Object Detection

Contents

1. Requirements: hardware
2. Requirements: software
3. Installation
4. Data preparation
5. Testing
6. Training

Requirements: hardware

• This code is GPU-only
• Tested on Ubuntu 18.04 with NVIDIA Tesla P100 and CUDA 9.0

Requirements: software

• Implemented by Python 3.6
• Core dependencies are listed in environment.yaml

1. Install Miniconda or Anaconda
2. Create an environment based on the environment file

   conda env create -f environment.yaml

3. Activate this environment

   conda activate saf

Installation

1. Clone the repository

   git clone https://github.com/SA-Framework/saf.pytorch
   cd saf.pytorch
   export SAF_ROOT=`pwd`

2. Compile the CUDA code

   cd $SAF_ROOT/libs
   sh make.sh

Data preparation

1. Create data folder

   mkdir -p $SAF_ROOT/data/VOC2007/
   cd $SAF_ROOT/data/VOC2007/

2. Download the training, validation, test data and VOCdevkit of PASCAL VOC 2007

   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
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar

3. Extract all these tars into one directory named VOCdevkit

   tar xvf VOCtrainval_06-Nov-2007.tar
   tar xvf VOCtest_06-Nov-2007.tar
   tar xvf VOCdevkit_18-May-2011.tar

4. Download the COCO format pascal annotations from here and put them into the annotations directory
5. Create symlinks to image files

   ln -s VOCdevkit/VOC2007/JPEGImages/

6. The directory structure should look like this

   $SAF_ROOT/data/VOC2007/
   $SAF_ROOT/data/VOC2007/annotations
   $SAF_ROOT/data/VOC2007/JPEGImages -> VOCdevkit/VOC2007/JPEGImages/
   $SAF_ROOT/data/VOC2007/VOCdevkit        

7. [Optional] Follow similar steps to get PASCAL VOC 2012
8. Download and put the precomputed proposals under $SAF_ROOT/data/precomputed_proposals/

Testing

Our trained models are available here. Put them under $SAF_ROOT/data/saf_models.

On trainval set (CorLoc)

python3 tools/test_net.py --cfg configs/baselines/vgg16_voc2007.yaml \
  --set MODEL.LOAD_IMAGENET_PRETRAINED_WEIGHTS False \
  --load_ckpt data/saf_models/voc2007.pth \
  --dataset voc2007trainval

On test set (mAP)

python3 tools/test_net.py --cfg configs/baselines/vgg16_voc2007.yaml \
  --set MODEL.LOAD_IMAGENET_PRETRAINED_WEIGHTS False \
  --load_ckpt data/saf_models/voc2007.pth \
  --dataset voc2007test

Note: Add --multi-gpu-testing if multiple gpus are available.

Training

Download the backbone VGG-16 model (pre-trained on ImageNet) and put it under $SAF_ROOT/data/pretrained_model/.

CUDA_VISIBLE_DEVICES=0 python3 tools/train_net_step.py --dataset voc2007 \
  --cfg configs/baselines/vgg16_voc2007.yaml --bs 1 --nw 4 --iter_size 4

Note: The current implementation only supports single-gpu training.