Low-light Object Detection

TL;DR

The paper tackles object detection in extremely low-light images by fusing predictions from multiple CO-DETR–based models trained on varied lighting data. It introduces an Instance-Adaptive Transformer (IAT) and leverages three datasets (dark, IAT-enhanced, and NUScene-augmented) along with test-time augmentations and an $IoU$-guided clustering strategy to select reliable detections. The three-model fusion approach improves robustness across lighting conditions and achieves the best reported performance (0.754) compared with individual models. This method has practical implications for surveillance and autonomous systems operating under challenging illumination, demonstrating the value of cross-condition training and fusion for low-light detection.

Abstract

In this competition we employed a model fusion approach to achieve object detection results close to those of real images. Our method is based on the CO-DETR model, which was trained on two sets of data: one containing images under dark conditions and another containing images enhanced with low-light conditions. We used various enhancement techniques on the test data to generate multiple sets of prediction results. Finally, we applied a clustering aggregation method guided by IoU thresholds to select the optimal results.

Figures (3)

• Figure 1: The example of competition dataset
• Figure 2: The structure of CO-DETR
• Figure 3: The process of fusion