Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation

TL;DR

This work tackles low-light object detection under zero-shot day-night domain adaptation by learning illumination-invariant reflectance representations. It introduces DAI-Net, which adds a reflectance decoder to a standard detector and enforces stability via an interchange-redecomposition-coherence mechanism, guided by pseudo ground-truth from a pretrained Retinex network. The approach combines mutual feature alignment, reconstruction-based decomposition losses, and a redecomposition coherence loss to achieve robust generalization without real low-light target data. Empirical results on ExDark, DARK FACE, and CODaN demonstrate strong dark-domain generalization and cross-task applicability, with comprehensive ablations validating each component.

Abstract

Detecting objects in low-light scenarios presents a persistent challenge, as detectors trained on well-lit data exhibit significant performance degradation on low-light data due to low visibility. Previous methods mitigate this issue by exploring image enhancement or object detection techniques with real low-light image datasets. However, the progress is impeded by the inherent difficulties about collecting and annotating low-light images. To address this challenge, we propose to boost low-light object detection with zero-shot day-night domain adaptation, which aims to generalize a detector from well-lit scenarios to low-light ones without requiring real low-light data. Revisiting Retinex theory in the low-level vision, we first design a reflectance representation learning module to learn Retinex-based illumination invariance in images with a carefully designed illumination invariance reinforcement strategy. Next, an interchange-redecomposition-coherence procedure is introduced to improve over the vanilla Retinex image decomposition process by performing two sequential image decompositions and introducing a redecomposition cohering loss. Extensive experiments on ExDark, DARK FACE, and CODaN datasets show strong low-light generalizability of our method. Our code is available at https://github.com/ZPDu/DAI-Net.

Figures (10)

• Figure 1: Left and right are respective well-lit and low-light datasets. FS: fully-supervised learning directly using low-light images and labels; DA: conventional domain adaptation with the access to low-light target domain data (and labels); ZSDA: zero-shot day-night domain adaptation without the access of domain data but only knowing the dark scenario as a prior.
• Figure 2: The structure overview of our method: we input a well-lit image and its corresponding synthesized low-light image as a pair into our framework. The pretrained Retinex decomposition net, shown as the gray block on the bottom, is frozen and only used during training to infer reflectance and illumination pseudo ground truth $\hat{R}^l,\hat{R}^n,\hat{L}^l,\hat{L}^n$ to supervise the reflectance decoder (Sec. \ref{['sec:ref']}). Specifically, illumination pseudo ground truth and the first-round reflectance predictions $R_1^l,R_1^n$ are forwarded into the proposed interchange-redecomposition-coherence procedure on the right block for reconstructing and redecomposing second-round reflectance predictions $R_2^l,R_2^n$, as well as calculating redecomposition cohering loss $\mathcal{L}_{rc}$ (Sec. \ref{['sec:redecomp']}). The inference is the same as the original detector that our method is built on, represented as green blocks.
• Figure 3: Feature visualization. The original image, backbone features from DSFD li2019cvpr, and our method are given in the three rows.
• Figure 4: Mean magnitude of feature channels before and after applying our method. We use discrete and continuous curves for the two distributions respectively only for the visualization purpose.
• Figure 5: Left: Qualitative examples of DSFD li2019cvpr, ours and fine-tuned ours (see Sec. \ref{['subsec:darkface']}) on DARK FACE. DARK FACE images are enhanced only for visualization. Right: Qualitative examples of MAET cui2021iccv, ours and the ground truth on ExDark (zoom in for details).