# Spectrum-inspired Low-light Image Translation for Saliency Detection

**Authors:** Kitty Varghese, Sudarshan Rajagopalan, Mohit Lamba, Kaushik Mitra  
**arXiv:** [2303.10145](https://arxiv.org/abs/2303.10145)  
**Date:** 2023-03-17  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2303.10145v1/

## Key points

- The paper presents a novel technique for improving the performance of saliency detection models under low-light conditions.
- Existing methods suffer from deteriorating performance due to the lack of well-lit images in training datasets.
- The proposed technique performs classical band-pass filtering in the Fourier space to transform well-lit images to low-light images and use them as a proxy for real low-light images.
- Unlike popular deep learning approaches, the proposed transformation is fast, simple, and easy to extend to other tasks such as low-light depth estimation.
- State-of-the-art saliency detection and depth estimation networks trained on their proxy low-light images perform significantly better on real low-light images than networks trained using existing strategies.
- Saliency detection models are important in various real-world applications such as robot navigation, satellite imagery, video summarization, foreground annotation, and action recognition.
- The proposed method's effectiveness is demonstrated through quantitative comparisons with existing techniques using standard evaluation metrics.
- This approach has significant potential for practical applications where accurate object detection is critical but limited by poor lighting conditions.

## Summary

The paper presents a novel technique for improving the performance of saliency detection models under low-light conditions. Existing methods suffer from deteriorating performance due to the lack of well-lit images in training datasets. To address this issue, the authors propose a technique that performs classical band-pass filtering in the Fourier space to transform well-lit images to low-light images and use them as a proxy for real low-light images. Unlike popular deep learning approaches that require learning thousands of parameters and enormous amounts of training data, the proposed transformation is fast, simple, and easy to extend to other tasks such as low-light depth estimation. The experiments conducted by the authors show that state-of-the-art saliency detection and depth estimation networks trained on their proxy low-light images perform significantly better on real low-light images than networks trained using existing strategies. The paper provides an overview of saliency detection models' importance in various real-world applications such as robot navigation, satellite imagery, video summarization, foreground annotation, and action recognition. The proposed method's effectiveness is demonstrated through quantitative comparisons with existing techniques using standard evaluation metrics. Overall, this paper presents an innovative solution to improve saliency detection models' performance under challenging lighting conditions without requiring extensive data collection or complex deep learning algorithms. This approach has significant potential for practical applications where accurate object detection is critical but limited by poor lighting conditions.