Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WaterFlow: Explicit Physics-Prior Rectified Flow for Underwater Saliency Mask Generation

Runting Li, Shijie Lian, Hua Li, Yutong Li, Wenhui Wu, Sam Kwong

TL;DR

WaterFlow addresses the challenges of underwater salient object detection by embedding explicit underwater physical priors into a rectified flow-based framework and adding temporal modeling. The approach combines a Temporal-Aware Conditional Aggregation Module (TACAM), an Underwater Physical Prior Module (UPPM), and an Underwater Task Object Flow (UTOF) built on conditional Rectified Flow to produce high-quality saliency masks. Key contributions include hierarchical encoding of physical priors, depth-guided priors during training, and the use of a continuous normalizing flow for efficient, conditional saliency generation, achieving state-of-the-art results on USOD10K and UFO-120 with improved inference speed. The work has practical significance for real-time underwater robotics and monitoring by enabling robust saliency detection under challenging underwater conditions.

Abstract

Underwater Salient Object Detection (USOD) faces significant challenges, including underwater image quality degradation and domain gaps. Existing methods tend to ignore the physical principles of underwater imaging or simply treat degradation phenomena in underwater images as interference factors that must be eliminated, failing to fully exploit the valuable information they contain. We propose WaterFlow, a rectified flow-based framework for underwater salient object detection that innovatively incorporates underwater physical imaging information as explicit priors directly into the network training process and introduces temporal dimension modeling, significantly enhancing the model's capability for salient object identification. On the USOD10K dataset, WaterFlow achieves a 0.072 gain in S_m, demonstrating the effectiveness and superiority of our method. The code will be published after the acceptance.

WaterFlow: Explicit Physics-Prior Rectified Flow for Underwater Saliency Mask Generation

TL;DR

WaterFlow addresses the challenges of underwater salient object detection by embedding explicit underwater physical priors into a rectified flow-based framework and adding temporal modeling. The approach combines a Temporal-Aware Conditional Aggregation Module (TACAM), an Underwater Physical Prior Module (UPPM), and an Underwater Task Object Flow (UTOF) built on conditional Rectified Flow to produce high-quality saliency masks. Key contributions include hierarchical encoding of physical priors, depth-guided priors during training, and the use of a continuous normalizing flow for efficient, conditional saliency generation, achieving state-of-the-art results on USOD10K and UFO-120 with improved inference speed. The work has practical significance for real-time underwater robotics and monitoring by enabling robust saliency detection under challenging underwater conditions.

Abstract

Underwater Salient Object Detection (USOD) faces significant challenges, including underwater image quality degradation and domain gaps. Existing methods tend to ignore the physical principles of underwater imaging or simply treat degradation phenomena in underwater images as interference factors that must be eliminated, failing to fully exploit the valuable information they contain. We propose WaterFlow, a rectified flow-based framework for underwater salient object detection that innovatively incorporates underwater physical imaging information as explicit priors directly into the network training process and introduces temporal dimension modeling, significantly enhancing the model's capability for salient object identification. On the USOD10K dataset, WaterFlow achieves a 0.072 gain in S_m, demonstrating the effectiveness and superiority of our method. The code will be published after the acceptance.
Paper Structure (11 sections, 7 equations, 3 figures, 2 tables)

This paper contains 11 sections, 7 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Overall Architecture
  5. Temporal-Aware Conditional Aggregation Module
  6. Underwater Physical Prior Module
  7. Underwater Task Object Flow
  8. Experiments
  9. Experiment Settings
  10. Evaluation and Ablation Study
  11. Conclusion

Figures (3)

  • Figure 1: The left panel shows the performance of our model on the USOD10K(L) and UFO-120(R) datasets, while the right panel compares our method against the previous SOTA approaches waterdiffusion. Red boxes highlight segmentation errors.
  • Figure 2: WaterFlow Architecture. The Temporal-Aware Conditional Aggregation Module (TACAM, defined in Sec. \ref{['sec:TACAM']}) and the Underwater Physical Prior Module (UPPM, defined in Sec. \ref{['sec:UPPM']}) jointly extract multi-scale features from the input images and masks, which are then fused to guide the downstream Underwater Task Object Flow (UTOF, defined in Sec. \ref{['sec:UTOF']}) in generating saliency segmentation masks.
  • Figure 3: Qualitative results of saliency detection for our method. In each group of images, the one on the left is the input image, the middle is the segmentation result generated by our model, and the one on the right is the ground truth (GT).