3D-USE: From Image-Level to Scene-Level Underwater Enhancement
summary
The gist
Underwater 3D reconstruction faithfully reproduces the color shifts and visibility loss of captured views…
In short
3D-USE reconstructs underwater scenes by learning a persistent, visibility-enhanced 3D representation from degraded views. It separates object appearance from water effects using MediumRBF and uses Appearance Transition Consensus (ATC) to guide scene enhancement. This two-stage framework produces consistent, high-quality enhanced renderings without needing a separate 2D model during inference.
Key concepts
- Medium Radial Basis Anchor Representation (MediumRBF)
- This component creates a medium-aware Gaussian scene. It uses isotropic radial basis anchors to interpolate medium properties in observer space. This allows the method to explicitly separate the object's appearance from the observer-dependent water effects, leading to a more stable geometry and explicit decomposition of the scene's medium.
- Appearance Transition Consensus (ATC)
- ATC learns transition priors by using paired 2D Underwater Image Enhancement (UIE) data. It consolidates these calibrated view-wise proposals into two types of guidance: scene-global targets and Gaussian-local targets. This ensures that the enhancement learned from individual views is coherently applied across the entire 3D scene.
- Underwater Bilateral Appearance Field (U-BAF)
- U-BAF realizes the learned guidance by performing coordinated modifications on both the Gaussian radiance and medium appearance. This process happens before underwater compositing. It allows for consistent novel-view rendering of enhanced scenes without requiring a 2D UIE model to be used at inference time.
Terminology used across episodes
This episode discusses
The paper
3D-USE: From Image-Level to Scene-Level Underwater Enhancement · Read on arXiv
Jieyu Yuan, Yuanlin Zhang, Jihong Li, Chunle Guo, Huimin Lu, Chongyi Li
College of Computer Science, Nankai University · NKIARI, Shenzhen Futian · School of Automation, Southeast University · Advanced Ocean Institute of Southeast University, Nantong
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "3D-USE: From Image-Level to Scene-Level Underwater Enhancement".
Jane: Here's the gist of the paper. Underwater 3D reconstruction faithfully reproduces the color shifts and visibility loss of captured views, while physical inversion may leave estimation errors in the recovered scene appearance The Core Problem Addressed Underwater 3D reconstruction faithfully reproduces the color shifts and visibility loss of captured views, while physical inversion may leave estimation errors in the recovered scene appearance.
Tom: Existing underwater 3DGS methods mainly obtain clear renderings through physical inversion by fitting a parametric underwater image-formation model and compensating for estimated scattering and attenuation.
Jane: However, scene radiance and medium parameters are jointly inferred from degraded observations, making their decomposition underdetermined.
Tom: Consequently, accurate reconstruction does not necessarily produce a reliable enhanced scene appearance.
Paper summary: Tom: So, wrapping up this discussion on "three dee-USE: From Image-Level to Scene-Level Underwater Enhancement <ref:2608.28020#pg1,3D-USE: From Image-Level to Scene-Level Underwater Enhancement>." This paper introduces a two-stage framework that first uses MediumRBF to build a medium-aware Gaussian scene and then employs Appearance Transition Consensus to define consistent enhancement targets. Jane The authors claim this allows for persistent, visibility-enhanced three dee scene representation from degraded multi-view underwater observations without needing paired enhanced three dee data <ref:2608.28020#pg1,persistent, visibility-enhanced 3D scene representation from degraded multi-view underwater observations>.
Meng: What this means practically is that if you have a set of underwater photos, you can use this method to get a consistent look across all views, even if the original photos were noisy or had bad visibility. Lalam It’s about making sure the enhancement isn't just an artifact of one specific view but something that stays the same whether you view it from angle one or angle twenty-six.
Lu: The method addresses two main challenges: getting a reliable scene representation that separates object appearance from water effects, and defining a consistent enhancement target when you don't have those perfect paired examples. Tom It’s about finding a way to get scene-level enhancement without relying on having perfectly enhanced three dee data for supervision <ref:2608.28020#pg1>.
Jane: The implication is that we can build three dee scenes that are inherently more robust to the visual degradation of underwater capture, which is a big step because currently, many methods struggle with view-to-view consistency in their enhancements <ref:2608.28020#pg1>. Tom So, it’s less about achieving one perfect image and more about achieving a consistent scene appearance across the whole set of views.
Meng: I think for deployment, this level of consistency is what matters most for applications where you need reliable three dee data from real-world underwater footage <ref:2608.28020#pg1>. Lalam The paper shows that combining global and local transition guidance leads to more uniform and natural results compared to using only one or the other type of supervision.
Conclusion: Tom: So, this paper is about something called three dee-USE, which is trying to take underwater photos and make them look better in three dee without having all that perfect paired data we usually need.
Jane: Right, so it’s moving beyond just fixing a single picture to making the whole scene look consistent across different views. The authors are focusing on this idea of learning a persistent, visibility-enhanced representation.
Lu: What they propose is a two-stage process where they first build this medium-aware Gaussian scene using something called MediumRBF to separate what the object looks like from how the water is messing with the image.
Meng: I’m wondering how much work that extra modeling does cost us in terms of training time and what breaks when we try to deploy it on, say, a real underwater drone feed.
Lalam: The core idea is that they use Appearance Transition Consensus, which learns from paired 2D images to create guidance for both the whole scene and the smaller Gaussian parts of the scene <ref:2608.28020#pg1>.
Tom: And then they realize this guidance is realized through this Underwater Bilateral Appearance Field, or U-BAF, which lets them modify both the color and the medium appearance for novel views without needing that extra 2D model during actual rendering <ref:2608.28020#pg1>.
Jane: It means they’re achieving consistency by learning how things *should* look across a set of photos rather than just correcting one photo at a time.
Tom: The optimization strategy involves minimizing a loss function that combines reconstruction error and depth accuracy in the first stage, and then matching those results to their learned targets in the second stage.
Lu: They found that explicit medium modeling really helps when you look at all these metrics, confirming that separating water effects from object appearance is genuinely beneficial for stability.
Meng: So they achieved better cross-view consistency, which is what we need for reliable underwater mapping or inspection data.
Lalam: They showed this method gets the lowest macro average on wLPIPS across a lot of real underwater scenes compared to other methods.
Jane: It seems like the main thing is that the enhanced appearance stays stable across different viewpoints rather than looking different every time you change your angle.
Tom: The big implication here for us is that we can get better three dee scenes from messy underwater data without needing perfect, expensive ground truth enhancement for every single view.
Jane: It shifts the focus from just restoration to building a representation that inherently knows how to handle the water distortion consistently.
Lu: And while they’re doing this scene-level enhancement, they also showed how combining global and local guidance helps balance things out—it avoids getting stuck either in a scene-wide color grade or just making small spatial fixes locally.
Meng: That's good because it means we don't have to choose between those two types of corrections; you get both at once.
Lalam: This work addresses the challenge of not having perfect paired data for supervision by using learned transition priors instead, which is a significant step forward in this area.
More episodes
- 2603.28595-Optimistic Actor-Critic with Parametric Policies for Linear Markov Decision Processes
- 2603.29765-Training-Free Dynamic Upcycling of Expert Language Models
- 2605.00778-Observable Performance Does Not Fully Reflect Adaptive System Organization: A Multi-Level Analysis of Gait Dynamics Under Occlusal Constraint
- 2605.06906-TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond
- 2605.09165-Sparse Layers are Critical to Scaling Looped Language Models
- 2605.09439-Inverse Design for Conditional Distribution Matching
- 2605.10313-Signature Approach for Contextual Bandits with Nonlinear and Path-dependent Rewards: Regret Bound and Statistical Inference
- 2605.10878-Minimum Neural Weight Norm = Kolmogorov Complexity
- 2605.11617-MIST: Reliable Streaming Decision Trees for Online Class-Incremental Learning via McDiarmid Bound
- 2605.15085-From Data to Action: Accelerating Refinery Optimization with AI