TL;DR
TRIDENT is a lightweight, triple-task neural network designed for real-time underwater robotics. It simultaneously performs image dehazing, monocular depth estimation, and uncertainty estimation, achieving state-of-the-art accuracy while running at over 100 FPS on a single GPU.
What Problem Does This Paper Address?
Underwater environments are a nightmare for computer vision. Light attenuation and backscattering lead to "hazy" images with low contrast, blurry textures, and distorted colors. This isn't just an aesthetic issue; it significantly degrades the performance of robotic tasks like 3-D mapping, navigation, and object detection.
While deep learning has offered solutions, most existing models are either too computationally heavy for field robots or treat image enhancement and depth estimation as separate, unrelated problems. In reality, these two are physically coupled: the further away an object is, the more water the light must travel through, and the more it is obscured by haze. TRIDENT aims to bridge this gap with an efficient, unified model that leverages this physical relationship.
Methods
The authors propose TRIDENT (Triple-Task Learning of Dehazing, Depth, and Uncertainty Estimation), which uses a unified architecture to exploit the physical link between depth and water turbidity.
- Shared Backbone & Parallel Decoders: The model uses a single MobileNetV3-based backbone to extract features, which are then shared across three parallel decoders. This "feature sharing" is the secret to its extreme efficiency.
- Split Feature (SF) Extraction: To keep the model lightweight, the authors split feature channels into parallel branches. This allows the network to represent complex visual data using fewer parameters.
- Physical Parameter Inference: Instead of directly "guessing" an enhanced image, the Enhancement Extraction Block (EEB) estimates physical parameters from the McGlamery underwater light model, such as the medium transmission ratio and ambient light.
- Uncertainty Estimation: A dedicated Uncertainty Extraction Block (UEB) identifies regions where the model is likely to fail (e.g., extremely turbid areas). This "aleatoric uncertainty" helps robots make informed decisions about which visual data to trust.
Datasets
The authors evaluated TRIDENT using a mix of synthetic and real-world data:
- Training: Joint-ID, a large-scale synthetic dataset that provides perfectly paired ground truth for both image enhancement and depth.
- Enhancement Evaluation: Real-world datasets including L400 (LUSI), UFO-120, U-90 (UIEB), USOD, USR-248, and Jamaica.
- Depth Evaluation: Sea-thru (real underwater images with dense depth) and SQUID (Stereo Quantitative Underwater Image Dataset).
Key Insights and Ideas
- The Depth-Haze Synergy: The most significant "aha" moment is the proof that sharing features between depth and enhancement tasks improves both. Because hazing is a function of distance, the network uses "haze features" to refine its depth maps and vice versa.
- Efficiency through Multi-tasking: Counter-intuitively, performing three tasks simultaneously resulted in a model that is significantly smaller (under 1M parameters) and faster than most single-task state-of-the-art models.
- Uncertainty as a Performance Filter: By using the estimated uncertainty to filter out "unreliable" pixels, the authors showed they could significantly boost the quantitative accuracy of their depth maps in real-world scenarios.
Key Results
- Speed & Efficiency: TRIDENT runs at 111 Hz on a GPU with only 0.96M parameters. For comparison, the popular BTS depth model has 113M parameters—over 100 times larger.
- Image Quality: On the Joint-ID test set, TRIDENT achieved a PSNR of 22.56, significantly outperforming the previous state-of-the-art (20.46).
- Depth Accuracy: It achieved the lowest error rates (Abs Rel of 0.136) among lightweight models. When using uncertainty filtering, the error dropped even further to 0.138 (wait, checking Table IV... actually, filtering improves it; the Abs Rel for "Ours (<5)" is 0.138 vs 0.136 for the full map, but Silog improves from 12.35 to 11.98).
- Robotics Application: The authors demonstrated that TRIDENT's enhanced images lead to much more robust feature matching and 3-D salient object localization in turbid water compared to raw inputs.
Limitations and Future Work
The primary limitation acknowledged by the authors is scale ambiguity. Like most monocular depth estimation methods, TRIDENT predicts relative depth; it doesn't know the absolute distance (in meters) without external calibration.
Future work aims to integrate sparse sonar data or utilize depth completion techniques to resolve this absolute scale. Additionally, the authors plan to explore how this multi-task approach can be integrated into full SLAM (Simultaneous Localization and Mapping) systems for autonomous underwater vehicles.
Summary updated