We propose 3DSmoothNet, a full workflow to match 3D point clouds with siamese deep learning architecture and fully convolutional layers using voxelized smoothed density value (SDV) representation. The latter is computed per interest aligned the local reference frame (LRF) achieve rotation invariance. Our compact, learned, invariant cloud descriptor achieves 94.9% average recall on 3DMatch benchmark data set, outperforming state-of-the-art by more than 20 percent points only 32 output...

We introduce PREDATOR, a model for pairwise point-cloud registration with deep attention to the overlap region. Different from previous work, our is specifically designed handle (also) pairs low overlap. Its key novelty an overlap-attention block early information exchange between latent encodings of two point clouds. In this way subsequent decoding representations into per-point features conditioned on respective other cloud, and thus can predict which points are not only salient, but also...

We present a novel, end-to-end learnable, multiview 3D point cloud registration algorithm. Registration of multiple scans typically follows two-stage pipeline: the initial pairwise alignment and globally consistent refinement. The former is often ambiguous due to low overlap neighboring clouds, symmetries repetitive scene parts. Therefore, latter global refinement aims at establishing cyclic consistency across helps in resolving cases. In this paper we propose, best our knowledge, first...

As several industries are moving towards modeling massive 3D virtual worlds, the need for content creation tools that can scale in terms of quantity, quality, and diversity is becoming evident. In our work, we aim to train performant generative models synthesize textured meshes which be directly consumed by rendering engines, thus immediately usable downstream applications. Prior works on either lack geometric details, limited mesh topology they produce, typically do not support textures, or...

Denoising diffusion models (DDMs) have shown promising results in 3D point cloud synthesis. To advance DDMs and make them useful for digital artists, we require (i) high generation quality, (ii) flexibility manipulation applications such as conditional synthesis shape interpolation, (iii) the ability to output smooth surfaces or meshes. this end, introduce hierarchical Latent Point Diffusion Model (LION) generation. LION is set up a variational autoencoder (VAE) with latent space that...

We propose a data-driven scene flow estimation algorithm exploiting the observation that many 3D scenes can be explained by collection of agents moving as rigid bodies. At core our method lies deep architecture able to reason at object-level considering in conjunction with other tasks. This object level abstraction enables us relax requirement for dense supervision simpler binary background segmentation mask and ego-motion annotations. Our mild requirements make well suited recently released...

We present a novel method for reconstructing 3D implicit surface from large-scale, sparse, and noisy point cloud. Our approach builds upon the recently introduced Neural Kernel Fields (NKF) [58] representation. It enjoys similar generalization capabilities to NKF, while simultaneously addressing its main limitations: (a) can scale large scenes through compactly supported kernel functions, which enable use of memory-efficient sparse linear solvers. (b) are robust noise, gradient fitting...

Reconstruction and intrinsic decomposition of scenes from captured imagery would enable many applications such as relighting virtual object insertion. Recent NeRF based methods achieve impressive fidelity 3D reconstruction, but bake the lighting shadows into radiance field, while mesh-based that facilitate through differentiable rendering have not yet scaled to complexity scale outdoor scenes. We present a novel inverse framework for large urban capable jointly reconstructing scene geometry,...

This work considers gradient-based mesh optimization, where we iteratively optimize for a 3D surface by representing it as the isosurface of scalar field, an increasingly common paradigm in applications including photogrammetry, generative modeling, and inverse physics. Existing implementations adapt classic extraction algorithms like Marching Cubes or Dual Contouring; these techniques were designed to extract meshes from fixed, known fields, optimization setting they lack degrees freedom...

We present Neural Fields for LiDAR (NFL), a method to optimise neural field scene representation from measurements, with the goal of synthesizing realistic scans novel viewpoints. NFL combines rendering power fields detailed, physically motivated model sensing process, thus enabling it accurately reproduce key sensor behaviors like beam divergence, secondary returns, and ray dropping. evaluate on synthetic real show that outperforms explicit reconstruct-then-simulate methods as well other...

We present Neural Kernel Fields: a novel method for reconstructing implicit 3D shapes based on learned kernel ridge regression. Our technique achieves state-of-the-art results when objects and large scenes from sparse oriented points, can reconstruct shape categories outside the training set with almost no drop in accuracy. The core insight of our approach is that methods are extremely effective chosen has an appropriate inductive bias. thus factor problem reconstruction into two parts: (1)...

Neural radiance fields achieve unprecedented quality for novel view synthesis, but their volumetric formulation remains expensive, requiring a huge number of samples to render high-resolution images. Volumetric encodings are essential represent fuzzy geometry such as foliage and hair, they well-suited stochastic optimization. Yet, many scenes ultimately consist largely solid surfaces which can be accurately rendered by single sample per pixel. Based on this insight, we propose neural that...

Abstract We propose a novel fully automated deformation analysis pipeline capable of estimating real 3D displacement vectors from point cloud data. Different the traditional methods that establish displacements based on proximity in Euclidean space, our approach estimates dense vector fields by searching for corresponding points across epochs space local feature descriptors. Due to this formulation, method is also sensitive motion and deformations occur parallel underlying surface. By...

We present SyNoRiM, a novel way to jointly register multiple non-rigid shapes by synchronizing the maps that relate learned functions defined on point clouds. Even though ability process is critical in various applications ranging from computer animation 3D digitization, literature still lacks robust and flexible framework match align collection of real, noisy scans observed under occlusions. Given set such clouds, our method first computes pairwise correspondences parameterized via...

Understanding and modeling lighting effects are fundamental tasks in computer vision graphics. Classic physically-based rendering (PBR) accurately simulates the light transport, but relies on precise scene representations--explicit 3D geometry, high-quality material properties, conditions--that often impractical to obtain real-world scenarios. Therefore, we introduce DiffusionRenderer, a neural approach that addresses dual problem of inverse forward within holistic framework. Leveraging...

We propose CaSPR, a method to learn object-centric Canonical Spatiotemporal Point Cloud Representations of dynamically moving or evolving objects. Our goal is enable information aggregation over time and the interrogation object state at any spatiotemporal neighborhood in past, observed not. Different from previous work, CaSPR learns representations that support spacetime continuity, are robust variable irregularly spacetime-sampled point clouds, generalize unseen instances. approach divides...

Abstract Lidar measurements and UAV photogrammetry provide high-resolution point clouds well suited for the investigation of slope deformations. Today, however, information contained in these is rarely fully exploited. This study shows three examples large-scale instabilities located Switzerland, which are actively monitored reasons hazard prevention. We used acquired by terrestrial laser scanning to (1) identify differences kinematic behaviour individual rock compartments; (2) highlight...

Particle-based representations of radiance fields such as 3D Gaussian Splatting have found great success for reconstructing and re-rendering complex scenes. Most existing methods render particles via rasterization, projecting them to screen space tiles processing in a sorted order. This work instead considers ray tracing the particles, building bounding volume hierarchy casting each pixel using high-performance GPU hardware. To efficiently handle large numbers semi-transparent we describe...

Abstract Areal deformation monitoring based on point clouds can be a very valuable alternative to the established point-based techniques, especially for of natural scenes. However, analysis approaches do not necessarily expose true 3D changes, because correspondence between points is typically naïvely. Recently, establish correspondences in feature space by using local descriptors that analyze geometric peculiarities neighborhood interest were proposed. resulting are noisy and contain large...

Abstract. The advantages of terrestrial laser scanning (TLS) for geodetic monitoring man-made and natural objects are not yet fully exploited. Herein we address one the open challenges by proposing feature-based methods identification corresponding points in point clouds two or more epochs. We propose a learned compact feature descriptor tailored outdoor scenes obtained using TLS. evaluate our method both on benchmark data set specially acquired dataset resembling simplified scenario where...

Humans naturally retain memories of permanent elements, while ephemeral moments often slip through the cracks memory. This selective retention is crucial for robotic perception, localization, and mapping. To endow robots with this capability, we introduce 3D Gaussian Mapping (3DGM), a self-supervised, camera-only offline mapping framework grounded in Splatting. 3DGM converts multitraverse RGB videos from same region into Gaussian-based environmental map concurrently performing 2D object...

We introduce OmniRe, a holistic approach for efficiently reconstructing high-fidelity dynamic urban scenes from on-device logs. Recent methods modeling driving sequences using neural radiance fields or Gaussian Splatting have demonstrated the potential of challenging scenes, but often overlook pedestrians and other non-vehicle actors, hindering complete pipeline scene reconstruction. To that end, we propose comprehensive 3DGS framework named allows accurate, full-length reconstruction...

We present a novel, end-to-end learnable, multiview 3D point cloud registration algorithm. Registration of multiple scans typically follows two-stage pipeline: the initial pairwise alignment and globally consistent refinement. The former is often ambiguous due to low overlap neighboring clouds, symmetries repetitive scene parts. Therefore, latter global refinement aims at establishing cyclic consistency across helps in resolving cases. In this paper we propose, best our knowledge, first...