Source-linked AI summary
Tri-DehazeGS: Scene--Medium Decoupled Gaussian Splatting with Transmittance-Aware Optimization
Kui Jiang, Yang Gu, Jiacheng Liu, Shiyu Liu, Youyu Chen, Hui Liu
TL;DR
Clean reconstruction from hazy multi-view images is hindered by scene–medium entanglement and weak supervision in low-transmittance regions. Tri-DehazeGS decouples clean Gaussian scenes from a shared tri-plane medium field and adds MD-TGC, improving clean novel-view reconstruction on real and synthetic haze benchmarks. Its restoration quality remains influenced by the accuracy of the estimated depth prior.
Problem
Hazy reconstruction must separate clean scene radiance from atmospheric effects while addressing weakened optimization supervision in low-transmittance regions.
Method
Tri-DehazeGS uses clean Gaussian primitives, an independent view-shared tri-plane medium field, physical scattering composition, and MD-TGC after medium stabilization.
Results
Tri-DehazeGS consistently improves clean novel-view synthesis quality over existing scattering-aware and restoration-assisted baselines on real and synthetic haze benchmarks.
Takeaways & Limitations
The paper frames physically degraded neural reconstruction as joint representation disentanglement and optimization rebalancing, rather than image-formation modeling alone.
Takeaways & Limitations
Restoration quality depends on estimated depth accuracy, with inaccurate or over-smoothed estimates causing blurred boundaries and lost fine details.
Abstract
from arXiv · showhide
Recovering clean 3D scenes from hazy multi-view images is challenging because haze attenuates scene radiance and introduces atmospheric scattering. Recent scattering-aware Gaussian Splatting methods introduce physical haze models into reconstruction, but they often apply degradation in image space or bind medium-related variables to Gaussian primitives, which can entangle clean scene radiance with atmospheric effects. Moreover, low-transmittance regions provide weakened supervision for Gaussian optimization, causing distant or dense-haze areas to be under-reconstructed. We argue that clean reconstruction under haze requires both scene--medium disentanglement and transmittance-aware optimization rebalancing. To this end, we propose Tri-DehazeGS, a scene--medium decoupled Gaussian Splatting framework. It represents the clean scene with Gaussian primitives, models the participating medium using an independent view-shared tri-plane field, and composes hazy observations through a physical scattering model. We further introduce Medium-Decoupled Transmittance Gradient Compensation (MD-TGC), which compensates haze-suppressed gradients after medium freezing without altering forward rendering. Experiments on real and synthetic haze benchmarks show that Tri-DehazeGS improves clean novel-view reconstruction. Code is available at https://github.com/aptx46/Tri-DehazeGS.
Introduction
Haze entangles scene radiance with atmospheric effects and weakens optimization in distant, dense-haze regions, so clean reconstruction requires decoupled scene–medium modeling and transmittance-aware supervision. Tri-DehazeGS addresses these issues with a shared tri-plane medium field and MD-TGC for clean novel-view reconstruction.
- Introduction: Standard 3DGS can absorb atmospheric attenuation and scattering into Gaussian colors, opacities, scales, or geometry instead of recovering the clean scene.
- Introduction: Low transmittance attenuates gradients to clean Gaussian primitives, leaving distant and heavily degraded structures under-optimized.This creates a supervision imbalance in addition to the representation entanglement caused by haze.
- Introduction: Existing approaches either apply degradation in image space without a shared 3D medium or bind medium variables to Gaussian primitives.These choices respectively lack cross-view medium consistency or couple scene content with medium effects.
- Introduction: Tri-DehazeGS independently represents the participating medium with a view-shared tri-plane field while retaining clean Gaussian scene primitives.The medium field organizes haze variables in shared 3D space and composes observations through an atmospheric scattering model.
- Introduction: MD-TGC freezes the stabilized medium and compensates transmittance-attenuated gradients without modifying physically based forward rendering.
Method
Tri-DehazeGS decouples clean Gaussian scene radiance from a view-shared participating-medium field, then rebalances haze-attenuated supervision through a two-stage optimization schedule.
- Scene–Medium Decoupled Rendering: Tri-DehazeGS jointly learns clean Gaussian scene primitives and an independent view-shared medium field, assigning scene radiance to one branch and attenuation or in-scattering to the other.The medium is represented continuously with a compact tri-plane field queried in a common three-dimensional coordinate system across views.
- Scene–Medium Decoupled Rendering: The medium branch predicts transmittance and scattering residuals, while the Gaussian branch renders clean radiance that can be recovered without the medium branch.The resulting decomposition provides a physically consistent separation of clean radiance, attenuation, and in-scattering.
- View-Shared Tri-Plane Medium Field: The tri-plane medium field predicts nonnegative extinction and direction-aware medium radiance, sharing atmospheric structure across views while retaining a compact explicit representation.Softplus enforces nonnegative extinction, and a learnable global scale separates spatial haze distribution from overall strength.
- Medium-Decoupled Transmittance Gradient Compensation: Transmittance attenuates gradients entering the clean Gaussian branch, leaving distant and dense-haze regions with insufficient updates despite potentially large reconstruction errors.The method treats transmittance as both an image-formation quantity and a visibility-aware supervision signal.
- Medium-Decoupled Transmittance Gradient Compensation: MD-TGC rescales only the backward gradient after freezing the medium field, preserving forward haze rendering while compensating transmittance-induced supervision attenuation.Freezing prevents the medium from artificially lowering transmittance to amplify Gaussian gradients.
- Two-Stage Optimization: The two-stage schedule first jointly optimizes scene and medium decomposition, then freezes the stabilized medium and refines the Gaussian scene with MD-TGC.Stage I also uses monocular pseudo-depth and a dark-channel transmittance prior; MD-TGC is disabled until the medium becomes reliable.
Experiments
Tri-DehazeGS consistently improves clean novel-view reconstruction across real and synthetic haze benchmarks, while ablations support decoupled, view-shared medium modeling and delayed transmittance-gradient compensation.
- Quantitative comparison: 3.35 dB PSNR and 0.056 SSIM improvements over 3D-UIR on RealX3D, with further PSNR gains over BiLaLoRA-GS and MarineSTD-GS on synthetic benchmarks.The gains are 2.28 dB on Mip-NeRF 360 foggy scenes and 0.93 dB on Fog-NeRF scenes.
- Qualitative comparison: Tri-DehazeGS reconstructs cleaner appearances while preserving object boundaries, distant structures, and local textures in qualitative comparisons.Existing methods show residual haze, color distortion, or weak distant-structure recovery.
- Scene–medium decoupling: Explicitly decoupling the medium from Gaussian primitives outperforms constant and coupled medium models, while direction-aware illumination improves reconstruction under non-uniform lighting.The constant model cannot represent spatially varying haze, and coupling atmospheric effects to Gaussian appearance remains inferior to explicit decoupling.
- Medium-field parameterization: The view-shared tri-plane medium representation balances compactness and expressiveness better than coordinate or Fourier MLPs, with voxel features achieving comparable performance.Coordinate and Fourier MLPs struggle with spatially varying haze, while qualitative results show the same trend.
- Transmittance-gradient compensation: MD-TGC achieves the best overall performance by freezing the medium before compensating transmittance-suppressed gradients, benefiting low-visibility regions most.Compensation from the beginning improves perceptual quality but reduces PSNR/SSIM because the evolving medium produces unstable signals; without TGC, low-transmittance regions receive insufficient gradients.
- Auxiliary supervision: Depth correlation provides the largest auxiliary-loss benefit, while depth smoothness and DCP act mainly as smaller stabilizers during scene–medium decomposition.The ablation indicates that depth priors help stabilize decomposition, although they are insufficient by themselves.
Limitations
Tri-DehazeGS depends on estimated depth priors, so inaccurate or over-smoothed depth can blur thin structures and low-texture boundaries and remove fine details.
- Limitations: Inaccurate or over-smoothed estimated depth can blur object boundaries and remove fine details, especially around thin structures and low-texture regions.The paper suggests more reliable depth priors or structure-aware constraints as possible improvements.
Conclusion
Tri-DehazeGS addresses haze-aware reconstruction through scene–medium separation and transmittance-aware optimization, improving clean novel-view synthesis on real and synthetic benchmarks. Its analyses attribute gains to both improved decomposition and optimization rebalancing, while identifying inaccurate depth priors as a failure case.
- Visibility-binned LPIPS analysis compares w/o TGC against MD-TGC across high, mid, and low transmittance ranges, reporting LPIPS decreases for the proposed method.The bins are defined by ˆT > 0.7, 0.4 < ˆT ≤0.7, and ˆT ≤0.4.
- A reported failure case is attributed to an inaccurate depth prior.
- The method separates scene radiance from participating-medium effects with a view-shared three-dimensional medium representation and restores haze-suppressed optimization signals using MD-TGC.
- Tri-DehazeGS consistently improves clean novel-view synthesis quality over existing scattering-aware and restoration-assisted baselines on real and synthetic haze benchmarks.
- Performance gains arise from both improved scene–medium decomposition and transmittance-aware optimization rebalancing.