Introduction

This repository contains the Torch implementation for ScaleNet: Guiding Object Proposal Generation in Supermarkets and Beyond at ICCV 2017. The code is built on DeepMask and SharpMask.

Citation

If you find ScaleNet useful in your research, please consider citing:

@inproceedings{ScaleNet,
   title = {ScaleNet: Guiding Object Proposal Generation in Supermarkets and Beyond},
   author = {Siyuan Qiao and Wei Shen and Weichao Qiu and Chenxi Liu and Alan Yuille},
   booktitle = {ICCV},
   year = {2017}
}

Get Started

1. Install the following packages for Torch: COCO API, image, tds, cjson, nnx, optim, inn, cutorch, cunn, cudnn
2. Clone this repository

SCALENET=/desired/absolute/path/to/scalenet/ # set absolute path as desired
git clone https://github.com/joe-siyuan-qiao/ScaleNet.git $SCALENET

3. Prepare environment

cd $SCALENET
mkdir -p data intermediate pretrained/scalenet pretrained/sharpmask

Download the pretrained ResNet-50 to $SCALENET/pretrained if you want to train ScaleNet or SharpMask. Move the downloaded MS COCO dataset to $SCALENET/data: $SCALENET/data/annotations, $SCALENET/data/train2014, $SCALENET/data/val2014.

Training and Evaluation

th trainScaleNet.lua # For ScaleNet
th train.lua # For DeepMask and SharpMask. Please see their repo for the training details

The trained models will be found in $SCALENET/exps. Move the trained models for ScaleNet and SharpMask into the corresponding folders $SCALENET/pretrained/scalenet and $SCALENET/pretrained/sharpmask. Our pretrained models can be found here: ScaleNet and SharpMask. Next, we can evaluate the models on MS COCO.

th evalCocoBbox.lua