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.
Introduction
Standard object detection models typically struggle in extremely low-light environments. As illumination decreases, image noise increases significantly, contrast disappears, and visual features that are usually distinct become buried in the background. Furthermore, in many low-light datasets, objects are often "stacked" or partially obscured, making it difficult for standard algorithms to delineate overlapping boundaries. This work aims to bridge the performance gap between high-quality, well-lit images and the noisy reality of low-light computer vision by accurately identifying and locating eight categories of everyday objects in extreme darkness.
Methodology
CO-DETR Architecture
The pipeline is built upon the CO-DETR (Collaborative Detection Transformer) architecture. Unlike standard DETR models, CO-DETR incorporates conventional object detector heads alongside the transformer-based head. This collaborative approach improves convergence by increasing the number of positive samples during training, which is crucial for learning meaningful features when the signal-to-noise ratio is low.
Multi-Stream Training Strategy
To maximize robustness, we trained three separate versions of the CO-DETR model:
- Dark Model: Trained on original, raw low-light images to capture features directly from the noisy input.
- Light Model: Trained on images pre-processed with an Instance-Adaptive Transformer (IAT), which dynamically adjusts brightness and contrast at the object level.
- Augmented Model: Trained on data augmented with the NUScenes dataset to improve generalizability across diverse scenes and lighting conditions.
Test-Time Augmentation (TTA) and Fusion
During the inference phase, we applied Test-Time Augmentation (TTA) by resizing images from 1200x800 to 1400x1000 and jittering HSV values. This generated multiple sets of predictions for each input. To resolve the "stacking" problem and merge these predictions, we utilized a clustering aggregation method guided by Intersection over Union (IoU) thresholds. This method selects the bounding box with the highest confidence score from each cluster, providing a cleaner and more accurate final output than standard Non-Maximum Suppression (NMS).
Experimental Results
The experimental results demonstrate a clear advantage for the fusion approach. While the baseline Dark-model achieved a result of 0.732, the IAT-enhanced Light-model improved this to 0.745. The final Fusion strategy, combining all streams and TTA, reached a peak performance of 0.754.
These results prove that aggregating predictions from models trained on different data distributions (raw dark vs. enhanced) is more effective than relying on any single model. The diversity of the training streams allows the system to compensate for the specific weaknesses of individual approaches.
Conclusion
This study presents a robust strategy for object detection in extreme low-light conditions. By leveraging the collaborative nature of CO-DETR and a multi-stream fusion pipeline, we achieved significant performance gains. Future work will focus on reducing the computational overhead of running multiple models and optimizing the system for real-time deployment on edge devices, where low-light performance is a critical safety requirement for autonomous systems.
Summary updated