GOI-YOLOv8 Grouping Offset and Isolated GiraffeDet Low-Light Target Detection

Mengqing Mei, Ziyu Zhou, Wei Liu, Zhiwei ỸeSensors 2024 · DOI

TL;DR

This paper introduces GOI-YOLOv8, a novel object detection model built upon YOLOv8, specifically designed for low-light conditions. By enhancing the feature fusion neck with a sampling offset mechanism and employing group convolution in the detection head, GOI-YOLOv8 significantly improves detection accuracy and efficiency in challenging low-light scenarios.

What Problem Does This Paper Address?

Object detection models, particularly state-of-the-art ones like YOLOv8, perform exceptionally well under good lighting conditions. However, their performance drastically degrades in low-light environments due to reduced visibility, noise, and color distortion. This limitation hinders their application in critical areas like autonomous driving at night, surveillance in dim conditions, and rescue operations. Existing solutions often involve separate low-light image enhancement steps, which can be computationally expensive and not real-time friendly, or they involve adapting existing detectors with minimal modifications, which often fall short of robust performance in truly dark scenes. The core challenge is to develop an end-to-end object detection system that can effectively extract and utilize information from low-light images without sacrificing speed.

Methods

The GOI-YOLOv8 model builds upon the YOLOv8 architecture with several key innovations:

  • RepGFPN Neck Enhancement: The authors adapt the repGFPN feature pyramid network, originally from DAMO-YOLO, to enhance the feature fusion capabilities of YOLOv8's neck. This involves increasing the hierarchical fusion levels to better integrate low-level and high-level spatial information crucial for low-light scenes.
  • Sampling Map Offset (DySample Inspiration): To address limitations in the repGFPN's upsampling module (specifically, potential loss of deep neck information due to fewer upsampling layers), they incorporate a sampling map offset mechanism inspired by DySample. This technique treats feature maps as continuous, using content-aware sampling points with controlled offsets to resample features, thereby improving the extraction of subtle details in low-light images. They also introduce a "static scope coefficient" to manage the sampling range and prevent artifacts.
  • Grouping Operation: To further refine feature extraction and mitigate issues like edge confusion common in low-light images, they employ group convolution in the detection head. This method divides feature maps into groups and processes them independently, effectively isolating interference information related to object edges. This also contributes to reducing computational complexity.

Datasets

The primary dataset used for training and evaluation in this paper is the ExDark dataset. This dataset is specifically designed for low-light object detection and image enhancement research. It contains images captured under various low-light conditions and includes annotations for objects and image levels. The authors divided ExDark into training, validation, and testing sets in an 8:1:1 ratio.

Key Insights and Ideas

  • Low-light detection is more than just brightness: The paper highlights that low-light detection isn't just about making images brighter; it's about effectively extracting and distinguishing features that are often obscured or distorted by poor illumination.
  • Feature fusion is critical, but needs adaptation: While feature pyramid networks are standard, their effectiveness in low-light scenarios can be limited. Enhancing them with deeper hierarchies and more sophisticated sampling (like offset sampling) is key.
  • Isolating noise is as important as finding features: The use of group convolution to specifically address edge confusion demonstrates that actively filtering out detrimental information is as crucial as enhancing useful signals in low-light detection.
  • Real-time performance is achievable with targeted optimizations: The authors show that by carefully modifying the neck and head of a powerful baseline like YOLOv8, significant improvements in both accuracy and speed can be achieved for low-light tasks.

Key Results

The proposed GOI-YOLO model demonstrates significant improvements on the ExDark dataset:

  • Accuracy: Achieved a competitive increase of 2.1% in mAP50 and 0.6% in mAP95 compared to the baseline YOLOv8-S/N. It also outperforms other leading models like YOLOv7-T, YOLOv5-S/N, YOLOV9-T, YOLOV10-N, and TOOD in terms of mAP50.
  • Efficiency: Reduced parameter count by 11% compared to YOLOv8 and decreased computational requirements by 28%.
  • Speed: Achieved the highest FPS among the compared models, indicating superior real-time performance.
  • Parameter Count: The model has a lower number of parameters (2.68m) than YOLOv8-S (3.01m) and YOLOv8-N (11.13m), making it more lightweight.

Qualitatively, visualizations show that GOI-YOLO effectively isolates the influence of edge information in low-light scenes, leading to more accurate bounding boxes and better capture of object details.

Limitations and Future Work

  • Dependence on Lighting: The authors acknowledge that accurate detection still requires some lighting and light source information, distinguishing low-light conditions from complete darkness.
  • Dataset Availability: They note that the scarcity of specialized low-light datasets is a general challenge, and while ExDark is valuable, more diverse low-light datasets could further improve model generalization.
  • Future Work: The paper implies that further exploration of dataset augmentation and potentially more advanced feature fusion techniques could lead to even greater improvements. The success of GOI-YOLO suggests that targeted architectural modifications can unlock significant potential for object detection in challenging visual environments.

Summary updated

Abstract

In the realm of computer vision, object detection holds significant importance and has demonstrated commendable performance across various scenarios. However, it typically requires favorable visibility conditions within the scene. Therefore, it is imperative to explore methodologies for conducting object detection under low-visibility circumstances. With its balanced combination of speed and accuracy, the state-of-the-art YOLOv8 framework has been recognized as one of the top algorithms for object detection, demonstrating outstanding performance results across a range of standard datasets. Nonetheless, current YOLO-series detection algorithms still face a significant challenge in detecting objects under low-light conditions. This is primarily due to the significant degradation in performance when detectors trained on illuminated data are applied to low-light datasets with limited visibility. To tackle this problem, we suggest a new model named Grouping Offset and Isolated GiraffeDet Target Detection-YOLO based on the YOLOv8 architecture. The proposed model demonstrates exceptional performance under low-light conditions. We employ the repGFPN feature pyramid network in the design of the feature fusion layer neck to enhance hierarchical fusion and deepen the integration of low-light information. Furthermore, we refine the repGFPN feature fusion layer by introducing a sampling map offset to address its limitations in terms of weight and efficiency, thereby better adapting it to real-time applications in low-light environments and emphasizing the potential features of such scenes. Additionally, we utilize group convolution to isolate interference information from detected object edges, resulting in improved detection performance and model efficiency. Experimental results demonstrate that our GOI-YOLO reduces the parameter count by 11% compared to YOLOv8 while decreasing computational requirements by 28%. This optimization significantly enhances real-time performance while achieving a competitive increase of 2.1% in Map50 and 0.6% in Map95 on the ExDark dataset.
BibTeX
@inproceedings{mei2024goi,
  author = {Mengqing Mei and Ziyu Zhou and Wei Liu and Zhiwei Ỹe},
  title = {{GOI-YOLOv8 Grouping Offset and Isolated GiraffeDet Low-Light Target Detection}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s24175787},
}
Figure