Search papers, labs, and topics across Lattice.
This paper introduces NemoSplat, a novel feed-forward 4D Gaussian Splatting framework specifically designed for dynamic underwater scene reconstruction from uncalibrated marine videos. By employing a Promptable Dynamic Disentangler and a Media-Aware Gaussian Predictor, the method effectively mitigates the challenges posed by optical attenuation and dynamic objects, achieving robust camera pose estimations and high-fidelity scene rendering. Experimental results on a newly created large-scale underwater dataset show that NemoSplat outperforms existing methods in both tracking accuracy and visual quality.
NemoSplat achieves state-of-the-art underwater scene reconstruction by isolating dynamic elements and enhancing visual fidelity in challenging aquatic environments.
Reconstructing photorealistic scenes in unconstrained underwater environments remains challenging due to severe media-induced light scattering and unpredictable dynamic objects. Recent feed-forward visual foundation models have demonstrated remarkable capabilities in generalized novel view synthesis and tracking. However, when directly applied to aquatic videos, optical attenuation and motion interference fatally corrupt their feature aggregation, leading to severe tracking and reconstruction failures. To overcome these limitations, we present NemoSplat, the first feed-forward 4D Gaussian Splatting framework tailored for media-aware dynamic reconstruction directly from uncalibrated marine videos. Beyond providing robust estimations of camera poses and dense scene depth, we devise a Promptable Dynamic Disentangler that utilizes a confidence-aware fusion strategy of learned dynamic probabilities and optional semantic text priors, effectively isolating massive transient entities. Furthermore, to counteract visual degradation, a Media-Aware Gaussian Predictor is formulated to jointly estimate intrinsic 3D Gaussian attributes alongside physical media parameters, rendering pristine scene appearance in a single forward pass. Additionally, we introduce a large-scale underwater dataset with massive dynamic elements to facilitate training and evaluation. Extensive experiments on our dataset demonstrate that NemoSplat achieves state-of-the-art tracking accuracy and high-fidelity rendering.
