AquaFeat+: an Underwater Vision Learning-based Enhancement Method for Object Detection, Classification, and Tracking

TL;DR

Underwater perception is hampered by color distortion and scattering, limiting automated tasks such as detection, classification, and tracking. The authors propose AquaFeat+, a plug-and-play, end-to-end trainable feature-centric enhancement module comprising a color correction stage, a multi-resolution Underwater-Feature Enhancement Network, a Global-Scale Attention Module, and an adaptive residual output. Evaluated on the FishTrack23 dataset with YOLO-based detectors and backbones, AquaFeat+ delivers state-of-the-art or competitive task performance, including a top F1-Score for detection and leading HOTA for tracking, while maintaining feasibility for real-time robotic systems. The approach demonstrates that task-oriented feature enhancement can outperform purely perceptual improvements and is adaptable to diverse architectures, suggesting practical benefits for underwater robotic perception. Future work will expand to depth estimation, semantic segmentation, and additional datasets and architectures to broaden applicability.

Abstract

Underwater video analysis is particularly challenging due to factors such as low lighting, color distortion, and turbidity, which compromise visual data quality and directly impact the performance of perception modules in robotic applications. This work proposes AquaFeat+, a plug-and-play pipeline designed to enhance features specifically for automated vision tasks, rather than for human perceptual quality. The architecture includes modules for color correction, hierarchical feature enhancement, and an adaptive residual output, which are trained end-to-end and guided directly by the loss function of the final application. Trained and evaluated in the FishTrack23 dataset, AquaFeat+ achieves significant improvements in object detection, classification, and tracking metrics, validating its effectiveness for enhancing perception tasks in underwater robotic applications.

Figures (4)

• Figure 1: Overview of the AquaFeat+ architecture. The pipeline begins with the color correction module, followed by the Underwater-Feature Enhancement Network (U-FEN) for feature extraction. The Global-Scale Attention Module (GSAM) aggregates global features and merges them with those from the original 1/8th-sized image to produce an enhanced image, which is then fed into the chosen task backbone.
• Figure 2: An overview of the Global-Scale Attention Module (GSAM), designed to enhance features by incorporating both global and multi-scale context. The module first processes the input along two distinct paths. One path captures long-range spatial dependencies using a global attention mechanism. The second path intelligently fuses features from different scales. The outputs of paths are combined with the original input feature map to produce a comprehensive and enriched representation.
• Figure 3: Qualitative comparison of model performance on the FishTrack23 dataset. (a) Bounding box predictions for the object detection task (YOLOv8m). (b) Species identification results for the classification task (YOLOv11s-cls). For the classification results, blue bounding boxes denote the ground truth, green boxes indicate a correct classification by the model, and red boxes signify a misclassification.
• Figure 4: Qualitative results of model performance across different methods in tracking in the FishTrack23 dataset. Blue boxes indicate the Ground truth, and red boxes the objects detected, while the colored lines inside them represent the movement in the ground truth and are verified by the methods.