I2-NeRF: Learning Neural Radiance Fields Under Physically-Grounded Media Interactions

TL;DR

I2-NeRF advances neural radiance fields by enforcing isometric and isotropic perception in media-rich environments. It introduces a general radiative formulation that unifies emission, absorption, and scattering via Beer-Lambert attenuation and models both objects and media as volumetric particles, coupled with a reverse-stratified upsampling strategy to preserve medium sampling around scenes. The approach yields physically plausible reconstructions across underwater, hazy, and low-light conditions and demonstrates depth estimation capabilities, competitive performance, and clear radiance–backscatter decomposition. This work contributes a practical framework for metric-aware 3D reconstruction in real-world degraded media, with potential for downstream physical-scene understanding and editing.

Abstract

Participating in efforts to endow generative AI with the 3D physical world perception, we propose I2-NeRF, a novel neural radiance field framework that enhances isometric and isotropic metric perception under media degradation. While existing NeRF models predominantly rely on object-centric sampling, I2-NeRF introduces a reverse-stratified upsampling strategy to achieve near-uniform sampling across 3D space, thereby preserving isometry. We further present a general radiative formulation for media degradation that unifies emission, absorption, and scattering into a particle model governed by the Beer-Lambert attenuation law. By composing the direct and media-induced in-scatter radiance, this formulation extends naturally to complex media environments such as underwater, haze, and even low-light scenes. By treating light propagation uniformly in both vertical and horizontal directions, I2-NeRF enables isotropic metric perception and can even estimate medium properties such as water depth. Experiments on real-world datasets demonstrate that our method significantly improves both reconstruction fidelity and physical plausibility compared to existing approaches.

Figures (17)

• Figure 1: The proposed general radiative formulation accommodates diverse media conditions, including underwater, hazy, and low-light environments. The scene shown here was captured in the Pacific Ocean near Okinawa. As a result of isometric and isotropic metric perception, the estimated depth closely matches the actual depth of the scene.
• Figure 2: Illustration of our particle model and pipeline. (a) depicts three types of particles in the media field that emit, absorb, or scatter radiance. (b) presents the media upsampling strategy designed to prevent media field collapse. (c) shows the model architecture, in which the object and the upsampled media points are processed by hash encoders followed by separate MLPs to predict density, color, and downwelling depth (in the case of scattering).
• Figure 3: Visualization of virtual absorption fields on the LOM dataset cui2024aleth: (a) restored clean radiance $J$; (b) illumination map $T_{\mathcal{P}}$ estimated unsupervisedly via the BCP; (c,d) Particle density profiles along the rays, corresponding to the green and red points illustrated in (a).
• Figure 4: Visualization of NVS and media decomposition on the SeaThru-NeRF dataset levy2023seathru. Our method precisely estimates the vertical water depth, which is inaccessible in baseline methods. Top-right corner displays the mean RGB values of the medium color $\mathrm{c}^{\rm med}$ from test views, alongside the real-world ranges simulated in akkaynak2017space across various Jerlov water types nils1968irradiance within depths 1-20 meters.
• Figure 5: Qualitative comparison of NVS on the LOM dataset cui2024aleth.