Learning from Pixel-Level Noisy Label : A New Perspective for Light Field Saliency Detection

This is a PyTorch implementation of our paper

Overall

Prerequisites

• Python 3.6.12

• Pytorch 1.2.0+

• torchvision 0.4.0+

Update

1. We released our code for joint training with depth and appearance, which is also our best performance model.

Usage

1. Clone the repository

git clone https://github.com/OLobbCode/NoiseLF.git
cd NoiseLF-code/

2. Download the datasets

Download the following datasets and unzip them.

• DUT-LF dataset，fetch code is ‘vecy’.
• HFUT dataset.
• LFSD dataset.
• The .txt file link for testing and training is here, code is 'joaa'.

3. Train

1. Set the c.DATA.TRAIN.ROOT and c.DATA.TRAIN.LIST path in config.py correctly.
2. We demo using VGG-19 as network backbone and train with a initial lr of 1e-5 for 30 epoches.
3. After training the result model will be stored under snapshot/exp_noiself folder.

Note：only support c.SOLVER.BATCH_SIZE=1

4. Test

For single dataset testing: you should set c.PHASE='test' in config.py, and set c.DATA.TEST.ROOT , c.DATA.TEST.LIST as yours.

python demo.py 

For evaluate :

python evaluate.py

All results saliency maps will be stored under 'Test/Out/exp_noiself_30/' folders in .png formats.

Thanks to MOLF.