PMNet: Coarse to Fine: Progressive and Multi-Task Learning for Salient Object Detection (ICPR 2020)

Pytorch implementation of "Coarse to Fine: Progressive and Multi-Task Learning for Salient Object Detection"

Qualitative Evaluation

Quantative Evaluation

Getting Started

Installation

• Clone this repository

git clone https://github.com/tiruss/PMNet.git

• You can install all the dependencies by

pip install -r requirements.txt

Download datasets

• Download training datasets from the link. DUTS-TR is our train dataset.

• Other datasets can download from the link [sal_eval_toolbox] Thank you for the awesome evaluation toolbox!

Run experiments from pretrained weight

• Download pretrained weight from the link. [Google Drive] [Baidu Drive] Baidu drive will be updated soon.

• If you want CRF for postprocessing, download PyCRF from the link.

• Run

python Test.py --weight [pretrained wieght] --input_dir [dataset name, default='DUTS-TE'] --crf [default='False']

• Pre-computed saliency maps can download from the link. [Google Drive] [Baidu Drive] Baidu drive will be updated soon.

Train from scratch

• DUTS-TR is our traning set for pair comparison

• First, you need make contour of dataset

python Make_Contour.py --data_dir [Dataset name ex)'DUTS-TR']

• Run

python Train.py --img_dir [DUTS-TR img dir] --gt_dir [DUTS-TR label dir] --contour_dir [DUTS-TR contour dir] --epoch --batch_size --gpus --down_scale