Radar2Shape: 3D Shape Reconstruction from High-Frequency Radar using Multiresolution Signed Distance Functions
Neel Sortur, Justin Goodwin, Purvik Patel, Luis Enrique Martinez Jr, Tzofi Klinghoffer, Rajmonda S. Caceres, Robin Walters · 2025
Read Radar2Shape: 3D Shape Reconstruction from High-Frequency Radar using Multiresolution Signed Distance Functions by Neel Sortur, Justin Goodwin, Purvik Patel, Luis Enrique Martinez Jr, Tzofi Klinghoffer, Rajmonda S. Caceres, Robin Walters (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-17484.txt.
Introduction
Neel Sortur, Justin Goodwin, Purvik Patel, Luis Enrique Martinez, Tzofi Klinghoffer, Rajmonda S Caceres, and Robin Walters present Radar2Shape. A system that rebuilds the three-dimensional shape of an object from high-frequency radar signals. Radar sees shapes in a tangled way, and real signals cover only a few viewing angles. Camera-based methods can build shapes from few views but fail when radar echoes are treated like photos.
Radar2Shape works in two stages. First it learns a tidy internal space of shape features at several levels of detail. Then a diffusion process grows a shape in that space guided by the radar signal and its frequencies. The aim is a method that handles arbitrary shapes under partial views, of interest for trade and aerospace sensing.
Catalog record
| CREATOR | Neel Sortur, Justin Goodwin, Purvik Patel, Luis Enrique Martinez Jr, Tzofi Klinghoffer, Rajmonda S. Caceres, Robin Walters |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-17484
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Radar2Shape: 3D Shape Reconstruction from High-Frequency Radar using Multiresolution Signed Distance Functions Neel Sortur, Justin Goodwin, Purvik Patel, Luis Enrique Martinez Jr, Tzofi Klinghoffer, Rajmonda S. Caceres, Robin Walters Abstract Determining the shape of 3D objects from high-frequency radar signals is analytically complex but critical for commercial and aerospace applications. Previous deep learning methods have been applied to radar modeling; however, they often fail to represent arbitrary shapes or have difficulty with real-world radar signals which are collected over limited viewing angles. Existing methods in optical 3D reconstruction can generate arbitrary shapes from limited camera views, but struggle when they naively treat the radar signal as a camera view. In this work, we present Radar2Shape, a denoising diffusion model that handles a partially observable radar signal for 3D reconstruction by correlating its frequencies with multiresolution shape features. Our method consists of a two-stage approach: first, Radar2Shape learns a regularized latent space with hierarchical resolutions of shape features, and second, it diffuses into this latent space by conditioning on the frequencies of the radar signal in an analogous coarse-to-fine manner. We demonstrate that Radar2Shape can successfully reconstruct arbitrary 3D shapes even from partially-observed radar signals, and we show robust generalization to two different simulation methods and real-world data. Additionally, we release two synthetic benchmark datasets to encourage future research in the high-frequency radar domain so that models like Radar2Shape can safely be adapted into real-world radar systems. Figure 1 : Overview. Radar2Shape solves the challenging task of 3D shape reconstruction from radar captured at limiting viewing angles. (a) Limited views cause self-occlusion, resulting in missing information in the measurement. (b) Our approach overcomes this ambiguity by using a data-driven diffusion prior with a novel coarse-to-fine refinement technique in signed distance function space. This method accurately generates occluded geometries based on partial radar measurements, leading to better performance than (c) existing domain-adapted methods that can fail with limited views and struggle even in full observability. † † DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited. This material is based upon work supported by the Under Secretary of War for Research and Engineering under Air Force Contract No. FA8702-15-D-0001 or FA8702-25-D-B002. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Under Secretary of War for Research and Engineering. © 2025 Massachusetts Institute of Technology. Delivered to the U.S. Government with Unlimited Rights, as defined in DFARS Part 252.227-7013 or 7014 (Feb 2014). Notwithstanding any copyright notice, U.S. Government rights in this work are defined by DFARS 252.227-7013 or DFARS 252.227-7014 as detailed above. Use of this work other than as specifically authorized by the U.S. Government may violate any copyrights that exist in this work. 1 Introduction Radar is a reliable sensing mechanism in adverse light and weather conditions with wide-ranging applications such as robotics [ 3 ] , autonomous driving [ 6 ] , and remote sensing [ 5 ] . It operates by transmitting radio waves and analyzing the response, echoes that return after striking an object. Types of radar are typically distinguished by their wavelength – longer wavelengths struggle to detect small objects, like raindrops or geologic particles, while shorter wavelengths (high-frequency radar) can detect a variety of sizes, but may be more noisy. In both regimes, geometrically characterizing and reconstructing an object from its radar response still presents a challenging inverse learning task. At long ranges, radar signals are often noisy and provide poor resolution [ 26 ] . Furthermore, radar sensors often do not fully observe an object at all viewing angles. This results in partial observability in the radar response that introduces uncertainties in the reconstruction process. In this work, we tackle this difficult problem of object reconstruction from long-range, high-frequency radar responses that are partially observed. Previous deep learning approaches to long-range high-frequency radar modeling have focused primarily on extracting high-level features for classification, segmentation, or pose estimation tasks [ 41 , 42 ] , which are still difficult open problems. However, many downstream tasks require full-shape reconstruction of an observed object, a higher dimensional and even more challenging problem. The computer vision community has separately developed models for 3D shape reconstruction – these models are typically conditioned on partial point clouds or multi-view images, and they often try to estimate camera intrinsic and extrinsic parameters. However, there are unique challenges when conditioning on a radar signal for 3D reconstruction instead of multi-view images. First, radar lacks analogous camera parameters to estimate because the observed shape does not correspond to simple geometric projections of the radar response. Second, the partial observability introduces high uncertainty. Much more of the object may be occluded from a radar’s line of sight compared to a camera’s single view. Additionally, the radar signal is spread across multiple frequencies, which can correspond to different resolutions of the object’s geometry. Previous methods have not focused on the difficult problem of full-shape reconstruction from long-range high-frequency radar signals or taken advantage of individual frequencies in the radar response. We propose Radar2Shape, a method that can reconstruct full 3D shapes from high-frequency radar responses by associating radar frequencies with shape resolutions. Our approach consists of two stages: 1) learning a multi-resolution, hierarchical latent space for 3D shapes, and 2) training a diffusion model to denoise in this space by conditioning on radar responses. The first stage uses a series of encodings and a VAE to learn a regularized latent space of vectors defining signed distance functions (SDFs) [ 12 ] . Instead of representing a shape with a single latent vector, we separate the latent vector into components that represent shape features at multiple resolutions (e.g. the thin structures in a shape versus its overall structure). This representation is created by projecting multi-resolution point cloud features onto triplanes of various spatial resolutions. These are then processed separately and combined as input into an SDF network. The second stage uses a Transformer backbone to predict the denoised sequence of latent shape vectors, iteratively from coarsest to finest resolution, conditioned with attention on embeddings of the corresponding radar resolution. Additionally, we incorporate a domain-relevant 2D shape prior [ 41 ] and propose a more efficient version of our method for this lower dimensional shape space by 1) encoding the shape space as a projection of our 2D shape parameterization and 2) using a U-Net to jointly encode the radar response and predict denoised latent shape vectors. Overall, we make four primary contributions: • We present Radar2Shape, a novel denoising diffusion model that reconstructs an object’s 3D geometry from partially-observable, high-frequency radar observations. • We show superior results compared to many 3D reconstruction models adapted to the radar domain and an existing competitive radar baseline. • We demonstrate a general method for learning multiresolution signed distance functions of 3D geometries. • We introduce the Manifold40-PO and Manifold40-PO-SBR benchmark datasets, the first public datasets of diverse meshes and simulated high-frequency radar responses for radar-based single object reconstruction. 2 Related Work 2.1 Diffusion Models Diffusion models [ 20 ] have emerged as a powerful generative model applicable in many scientific domains ranging from bioinformatics [ 18 ] to climate science [ 30 ,