Saliency Detection in 360° Videos

This figure indicates how spherical crown kernel changes on sphere and projected panorama from north pole to south pole with angle interval equaling π=4. The first row is the region of the spherical crown kernel on sphere. The second row shows the region of spherical crown kernel on the projected panorama. The third row shows sampling grid corresponding to each kernel location. Red curve represents θ sampling grid and blue curve represents φ sampling grid.

Introduction

This repo contains the codes that used in paper Saliency Detection in 360° Videos by Ziheng Zhang, Yanyu Xu, Jingyi Yu and Shenghua Gao.

Requirements

• Python 3.6 is required.
• Module Dependencies:
  • 0.3.0 <= torch <= 0.3.1
  • numpy >= 1.13

File structure

- test.py
  purpose: Provides a simple test model that uses spherical convolution.
- spherical_unet.py
  purpose: Provides the implementation of Spherical U-Net that we used in our paper.
- train.py
  purpose: Provides training codes for Spherical U-Net.
- data.py
  purpose: Provides the dataloader for our dataset to train Spherical U-Net.
- sconv
  - functional
    - common.py
      Purpose: Contains some helper functions used in sphercal convolution.
    - sconv.py
      Purpose: Provides the spherical convolution function for Pytorch.
    - spad.py
      Purpose: Provides the spherical pooling function for Pytorch.
  - module
    - sconv.py
      Purpose: Provides the spherical convolution module for Pytorch.
    - smse.py
      Purpose: Provides the spherical mean-square loss module for Pytorch.
    - spad.py
      Purpose: Provides the spherical padding module for Pytorch.
    - spool.py
      Purpose: Provides the spherical pooling module for Pytorch.

Usage

The spherical convolution is written in pure python with pytorch, so that no compiling proceedure is needed. One can just pull and run all the codes in this repo. We currently provide a sample model in test.py that uses spherical convolution layers. The model and checkpoint that used in original paper will be released later.

Known issues

• The process of determining the kernel area at different θ locations is unstable, which will cause output feature maps contain some nan values. However, this bug seems to have minor effects during training and testing. We will try to fix it later.

Dataset

You can download our dataset [Baidu Pan] (code:p0a5), which consists of 104 videos. There are 12 zip files, and train/test index. After downloading these zips, unzip them together. There are 104 files and 'vinfo.pkl'. Each file consists of .jpg (RGB image) and .npy (ground truth heatmaps). The 'pkl' file consists of the original groud truth gaze points of the observers.

TODO

• [x] Release core functions and modules
• [x] Release training code for saliency detection
• [ ] Resolve the math unstability when calculating the kernel area at different θ locations
• [ ] Rewrite spherical convolution for torch 0.4+

License

This project is released under the MIT License (refer to the LICENSE file for details).

Citing

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

    @InProceedings{Zhang_2018_ECCV,
        author = {Zhang, Ziheng and Xu, Yanyu and Yu, Jingyi and Gao, Shenghua},
        title = {Saliency Detection in 360° Videos},
        booktitle = {The European Conference on Computer Vision (ECCV)},
        month = {September},
        year = {2018}
    }