WiSER: A Wireless Scene Encoder

WiSER: A Wireless Scene Encoder for Geometry-Grounded Multi-View Wireless Prediction

A sparse 3D scene representation that jointly supports dense radiomap prediction and multipath channel impulse response (CIR) tap-set prediction.

University of California Santa Cruz

Paper coming soon Code coming soon Data coming soon Model coming soon

Abstract

Indoor wireless propagation is governed by the interaction among 3D scene geometry, radio-material properties, and transmitter and receiver configuration. Most learning-based site-specific prediction methods focus on a single wireless representation, such as radiomap estimation or CIR prediction, and therefore do not explicitly exploit the propagation structure shared across heterogeneous wireless views.

WiSER maps a sparse voxel representation of an indoor scene and a transmitter location into a transmitter-conditioned sparse 3D scene memory. This shared memory is queried by two structure-aware decoders: a ray-corridor decoder for dense receiver-plane path-gain prediction and a Detection Transformer-style set decoder for variable-cardinality delay and power tap prediction.

We train and evaluate WiSER with a co-registered indoor scene--wireless dataset generated from ScanNet++ scenes and Sionna Ray Tracing. The dataset aligns sparse voxel inputs, dense radiomap labels, and unordered multipath CIR tap sets under a common coordinate frame and propagation configuration.

Key Idea

Shared Wireless Scene Memory

WiSER encodes sparse 3D scene voxels into a transmitter-conditioned memory that can be reused by multiple wireless prediction views.

Ray-Corridor Radiomap Decoding

The radiomap branch gathers receiver-specific scene tokens near the transmitter--receiver corridor to decode dense path-gain fields.

DETR-Style CIR Set Prediction

The CIR branch predicts unordered multipath delay--power taps with learnable path queries and Hungarian matching.

Method Overview

WiSER first builds a transmitter-conditioned sparse 3D scene memory. A ray-corridor radiomap decoder predicts dense receiver-plane path gain, while a CIR set decoder predicts variable-cardinality delay and power taps for a specific transmitter--receiver link.

Ray-corridor feature gathering

For each receiver query, WiSER selects a compact set of scene voxels near the transmitter--receiver segment and endpoint neighborhoods. This gives the radiomap decoder access to likely blockers, openings, reflectors, and nearby scattering structures without dense attention over the full indoor volume.

Co-Registered Scene--Wireless Dataset

WiSER is trained with a co-registered dataset pipeline that converts indoor 3D scenes into both sparse voxel inputs for learning and Sionna-compatible radio scenes for ray-tracing supervision.

The same coordinate frame produces aligned dense radiomap labels and path-level CIR labels. This makes it possible to study a single learned scene representation across coverage-level and path-level wireless views.

100 training scenes

10 cm voxel size

2 wireless views

The dataset pipeline aligns sparse voxel scenes, Sionna material scenes, dense radiomaps, and multipath CIR tap sets.

Results

Radiomap Prediction

3.834 dB

MAE on evaluated radiomap cases

WiSER improves over scene-specific NeRF2 and RF-3DGS baselines while being trained once across multiple scenes.

Multipath CIR Prediction

5.89 dB

matched peak-power MAE

The CIR decoder reduces matched peak-power and delay errors over geometry-only and 3D CNN reference baselines.

Shared Representation

0.61 ns

matched delay MAE

Results support the central claim that a sparse 3D scene memory can serve both dense field-level and sparse path-level prediction.

Qualitative radiomap examples compare ground truth, WiSER, NeRF2, RF-3DGS, and radiomap-head ablations under the same dB color scale.

Qualitative CIR examples show matched predicted taps against ground-truth delay--power taps in the delay/power plane.

Task	Method	Primary Metric	Additional Metrics
Radiomap	WiSER	3.834 dB MAE	5.500 dB RMSE, 26.78 dB PSNR
Radiomap	RF-3DGS	4.585 dB MAE	6.281 dB RMSE, 25.62 dB PSNR
CIR	WiSER	5.89 dB peak-power MAE	0.61 ns delay MAE, 0.477 count accuracy
CIR	3D CNN	11.50 dB peak-power MAE	1.50 ns delay MAE, 0.407 count accuracy

Release Status

We are preparing the public release of the WiSER codebase, processed dataset, and model checkpoint. The current project page is a preview; public links will be added after the corresponding repositories and archives are finalized.

Code: coming soon Dataset: coming soon Model: coming soon Paper/arXiv: coming soon