[ECCV2024] Watching it in Dark: A Target-aware Representation Learning Framework for High-Level Vision Tasks in Low Illumination

Yunan Li Yihao Zhang Shoude Li Long Tian Dou Quan Chaoneng Li Qiguang Miao
Xidian University; Xi'an Key Laboratory of Big Data and Intelligent Vision

This is the official implementaion of paper Watching it in Dark: A Target-aware Representation Learning Framework for High-Level Vision Tasks in Low Illumination, which is accepted in ECCV 2024. In this paper, we propose a target-aware representation learning framework to enhance high-level task performance in low-light conditions. We achieve bi-directional domain alignment using image appearance and semantic features, and introduce a target highlighting strategy with saliency mechanisms and Temporal Gaussian Mixture Model to emphasize task-relevant targets. Additionally, wedesign a mask token-based representation learning scheme to learn a more robust target-aware feature. Our framework is validated through extensive experiments on CODaN, ExDark, and ARID datasets, demonstrating effectiveness in classification, detection, and action recognition tasks.

👀TODO

First Release.
Release Code of Image Classification.
Release Code of Object Detection.
Release Code of Action Recognition.

🌏 Pipeline of WiiD

📚 Dataset

Data file name
Common Objects Day and Night (CODaN)
Exclusively Dark Image Dataset (ExDark)
normal light data of action recognition(CVPR'22 UG2 challenge)
low light data of action recognition(ARID dataset)

🐒 Model Zoo

will release

💻 Code