You Only Look Around: Learning Illumination Invariant Feature for Low-light Object Detection

TL;DR

This paper proposes to learn illumination-invariant features through the Lambertian image formation model, and introduces a novel module dedicated to the extraction of illumination-invariant features from low-light images, which can be easily integrated into existing object detection frameworks.

Abstract

In this paper, we introduce YOLA, a novel framework for object detection in low-light scenarios. Unlike previous works, we propose to tackle this challenging problem from the perspective of feature learning. Specifically, we propose to learn illumination-invariant features through the Lambertian image formation model. We observe that, under the Lambertian assumption, it is feasible to approximate illumination-invariant feature maps by exploiting the interrelationships between neighboring color channels and spatially adjacent pixels. By incorporating additional constraints, these relationships can be characterized in the form of convolutional kernels, which can be trained in a detection-driven manner within a network. Towards this end, we introduce a novel module dedicated to the extraction of illumination-invariant features from low-light images, which can be easily integrated into existing object detection frameworks. Our empirical findings reveal significant improvements in low-light object detection tasks, as well as promising results in both well-lit and over-lit scenarios. Code is available at \url{https://github.com/MingboHong/YOLA}.

Figures (8)

• Figure 1: (a): The base detector failed to recognize objects. (b, c) However, when IIM is employed with a simple edge feature, the object is identified. (d, e) Furthermore, the full IIM utilizes a task-driven learnable kernel to extract illumination-invariant features that are richer and more suitable for the detection task than simple edge features.
• Figure 2: The overall pipeline of YOLA.YOLA extracts illumination-invariant features via IIM and integrates them with original images by leveraging a fuse convolution block for the subsequent detector.
• Figure 3: Qualitative comparisons of TOOD detector on both ExDark and $UG^{2}+$DARK FACE dataset, where the top 2 rows visualize the detection results from ExDark, and the bottom 2 rows show the results from $UG^{2}+$DARK FACE. The images are being replaced with enhanced images generated by LLIE or low-light object methods. Red dash boxes highlight the inconspicuous cases. Zoom in red dash boxes for the best view.
• Figure 4: Visualization of the features (columns 2 and 4) generated by IIM-Edge and IIM(kernels are normalized for better visibility, we average the features across the channel dimensions and then conduct spatial normalization), along with detection results (columns 1 and 3). Best viewed by zooming in.
• Figure 5: Illustration of a trivial case (a), and visualization of performing $3\times3$ mean filtering on the kernel weights guided by with (b) and without (c) II loss.