beckachuu / frequency

Research frequency properties in images (WIP)

Frequency Filter

Description

This project aims to study the intriguing characteristics of images when divided into high-frequency and low-frequency components.

Table of Contents

• Getting Started

  • Prerequisites
  • Installation
  • Data Preparation
  • Input Configuration

• Run experiment

  • Run
  • View Result

• References

Getting started

Prerequisites

• Python

Installation

git clone https://github.com/beckachuu/FrequencyFilter.git
cd FrequencyFilter
pip install -r requirements.txt

Note: You may optionally wish to create a Python Virtual Environment to prevent conflicts with your system's Python environment.

Data Preparation

Image:

• COCO 2017 Val images

Groundtruth data:

• COCO 2017 Val annotations

Input Configuration

General input required arguments are in config.ini.

Run analysis and experiments

Run

Generate images

• Run analyze: separate high and low frequency components of input images

python analyze_freq.py

• Run experiments: run a chosen experiment

python run_exp.py

Detect objects

• Detect objects for experiment results

python detect.py

• Plot detected bboxes for experiment results

python plot_exp_demo.py

Evaluate

• Evaluate mean Average Precision (mAP) for experiment results

python eval.py

Extra

• Experiment with kernel: analyze magnitude and phase of different types of kernels in Fourier domain

python kernel_exp.py

View Result

• \output\{Your input folder name}\analyze\: analyze results
• \output\{Your input folder name}\EXP_{#}\: experiments results
• Detected images: open folder \detects\ inside each result folder

References

• The code obtaining low and high frequency component is partially obtained from the repository of Wang et al.

@article{Wang2019HighFrequencyCH,
  title={High-Frequency Component Helps Explain the Generalization of Convolutional Neural Networks},
  author={Haohan Wang and Xindi Wu and Pengcheng Yin and Eric P. Xing},
  journal={2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2019},
  pages={8681-8691},
  url={https://api.semanticscholar.org/CorpusID:173188317}
}

This project welcomes contributions and suggestions.