Source-linked AI summary
PE-YOLO: Pyramid Enhancement Network for Dark Object Detection
Xiangchen Yin, Zhenda Yu, Zetao Fei, Wenjun Lv, Xin Gao
TL;DR
Dark conditions remain a challenge for object detection. PE-YOLO combines a pyramid enhancement network with YOLOv3 in an end-to-end framework, and reports favorable ExDark results, including 78.0% mAP and 53.6 FPS.
Problem
Poor lighting reduces image quality and affects detector performance, while most existing detectors are studied on high-quality images under normal conditions.
Source: §1 · p. 2Method
PE-YOLO joins PENet with YOLOv3, using pyramid decomposition, DPM and LEF for component enhancement and end-to-end training with normal detection loss.
Source: §1 · p. 3Results
78.0% mAP and 53.6 FPS are reported on ExDark, with PE-YOLO compared favorably against dark detectors and low-light enhancement models.
Source: §1 · p. 3Takeaways & Limitations
The authors report that PE-YOLO can adapt object detection to different low-light conditions.
Source: §1 · p. 3
Abstract
from arXiv · showhide
Current object detection models have achieved good results on many benchmark datasets, detecting objects in dark conditions remains a large challenge. To address this issue, we propose a pyramid enhanced network (PENet) and joint it with YOLOv3 to build a dark object detection framework named PE-YOLO. Firstly, PENet decomposes the image into four components of different resolutions using the Laplacian pyramid. Specifically we propose a detail processing module (DPM) to enhance the detail of images, which consists of context branch and edge branch. In addition, we propose a low-frequency enhancement filter (LEF) to capture low-frequency semantics and prevent high-frequency noise. PE-YOLO adopts an end-to-end joint training approach and only uses normal detection loss to simplify the training process. We conduct experiments on the low-light object detection dataset ExDark to demonstrate the effectiveness of ours. The results indicate that compared with other dark detectors and low-light enhancement models, PE-YOLO achieves the advanced results, achieving 78.0% in mAP and 53.6 in FPS, respectively, which can adapt to object detection under different low-light conditions. The code is available at https://github.com/XiangchenYin/PE-YOLO.
1 Introduction
Dark lighting degrades image quality and challenges object detection. PE-YOLO combines pyramid-based image enhancement with YOLOv3 and reports strong ExDark detection accuracy and speed.
1 Introduction
- Object detection models are commonly evaluated on benchmark datasets, but most existing detectors are studied on high-quality images under normal conditions.Poor lighting can lower image quality and affect detector performance.Source: §1 · pp. 1–2
- 78.0% mAP and 53.6 FPS are reported for PE-YOLO, which the authors compare favorably with dark detectors and low-light enhancement models.The introduction reports these results on ExDark.Source: §1 · p. 3
- PENet decomposes images into components at different resolutions and enhances them with DPM and LEF.DPM combines context and edge branches; LEF captures low-frequency semantics and prevents high-frequency noise.Source: §1 · p. 3
- PE-YOLO joins PENet with YOLOv3 for end-to-end dark object detection, trained using only normal detection loss.The approach does not require clear image ground truth.Source: §1 · p. 3
2 Related Work
Related work covers general object detectors, low-light image enhancement, and detection methods for adverse conditions, including dark scenes.
2.1 Object Detection
- YOLOv3 uses DarkNet-53 for feature extraction and adopts multi-scale feature fusion inspired by feature pyramid networks.The section also describes Faster RCNN, Cascade RCNN, and anchor-free detectors.Source: §2.1 · p. 3
2.2 Low-light Enhancement
- Low-light enhancement methods restore image details or correct color distortion, but their complexity can substantially affect detector real-time performance.The section mentions Kind, Zero DCE, MBLLEN, and IAT as examples.Source: §2.2 · p. 4
2.3 Object Detection in Adverse Condition
- Adverse-condition detection approaches include unsupervised domain adaptation, differentiable image processing, and methods developed for dark scenes.Examples include IA-YOLO, GDIP-YOLO, DENet, and MAET.Source: §2.3 · p. 4
3 Method
PE-YOLO combines PENet with YOLOv3 for dark object detection. PENet enhances multiscale Laplacian-pyramid components using detail and low-frequency processing modules.
3.1 Overview of PE-YOLO
- PENet decomposes images into Laplacian-pyramid components at different scales, enhances them, and reverses the decomposition to restore a high-resolution image.The components represent information lost during downsampling; the pyramid emphasizes global information at some scales and local detail at others.Source: §3.1 · p. 5
3.2 Detail Enhancement
- DPM enhances pyramid components through a context branch that captures remote dependencies and an edge branch that uses horizontal and vertical Sobel gradients to enhance texture.The context branch uses residual blocks to change channels from 3 to 32 and back to 3; the edge branch uses residuals to support information flow.Source: §3.2 · pp. 6–7
3.3 Low-Frequency Enhancement Filter
- LEF captures low-frequency information in each component, using different average-pooling scales to filter features before upsampling them to the original size.The components are transformed from 3 to 32 channels, split into four parts, and processed with pooling sizes 1×1, 2×2, 3×3, and 6×6.Source: §3.3 · p. 7
4 Experiments
Experiments on ExDark compare PE-YOLO with low-light enhancement models and dark detectors, then assess its components through ablation. The reported comparisons favor PE-YOLO, and the ablation attributes successive mAP gains to the context branch, edge branch, and LEF.
4.1 Dataset and Implementation Details
- On ExDark, the study uses an 80/20 train/test split, 608 × 608 images, and mAP and FPS to evaluate detection accuracy and speed.Training runs for 30 epochs on one RTX 3090 GPU.Source: §4.1 · pp. 7–8
4.2 Experimental Results
- PE-YOLO’s mAP is 1.2% higher than MBLLEN’s and 1.1% higher than Zero-DCE’s, achieving the best results among the compared enhancement models.The comparison uses YOLOv3 on enhanced images and an mAP IoU threshold of 0.5.Source: §4.2 · p. 8, Table 1 · p. 9
- PE-YOLO’s mAP is 0.7% higher than DENet’s and 0.2% higher than LOL-pretrained IAT-YOLO’s, and its FPS is almost the highest.The authors conclude PE-YOLO is more suitable for dark-condition object detection.Source: §4.2 · p. 9, Table 2 · p. 10
4.3 Ablation Study
- Adding the context branch raises mAP from 76.4% to 77.0%, adding the edge branch raises it from 77.0% to 77.6%, and adding LEF raises it from 77.6% to 78.0%.The authors associate these gains, respectively, with capturing remote dependencies, enhancing component texture and image details, and capturing low-frequency information; the full model decreases FPS by 0.4.Source: §4.3 · pp. 9–10, Table 3 · p. 10
5 Conclusion
PE-YOLO combines PENet with YOLOv3 for end-to-end dark object detection. On ExDark, it achieves the best results compared with low-light enhancement models and dark detectors, while the authors identify further evaluation and lightweight performance as future work.
- On ExDark, PE-YOLO achieves the best results compared with low-light enhancement models and dark detectors.The conclusion also says it can effectively detect objects in dark conditions.Source: §5 · p. 11
- PENet decomposes images into four resolutions for detail restoration and potential-information capture, and PE-YOLO trains end to end without an additional loss function.It combines PENet with YOLOv3 and uses DPM and LEF to enhance the components.Source: §5 · p. 11
- The model remains to be studied with more detectors and to improve performance while maintaining a lightweight design.Source: §5 · p. 11