Anonymous Project Page

PRISM: Human Point Cloud Reconstruction
via Skeleton-Guided Diffusion
from mmWave Radar

Introducing MIST — the first mmWave dataset with a physical barrier between sensor and subject — and PRISM, a skeleton-guided latent diffusion framework that reconstructs LiDAR-quality human point clouds from sparse radar returns under three controlled occlusion levels.

Anonymous Authors · Anonymous Institution

📄 Paper 💻 Code 🗂️ Dataset 📊 Results

Abstract

From sparse radar signals
to complete human shapes

mmWave FMCW radar offers all-weather, non-contact, privacy-preserving human sensing, yet its inherent sparsity severely limits downstream body analysis and no existing dataset provides a paired quantitative benchmark for dense human point cloud reconstruction under occlusion.

We present MIST, the first mmWave human point cloud dataset with a physical barrier between sensor and subject. MIST pairs clear-view and through-obstacle captures of the same subjects performing the same actions, with synchronized LiDAR as 3D ground truth and a three-level occlusion design converting attenuation severity into a controlled variable — spanning 8 subjects, 30 action classes, and 640k paired frames.

We also propose PRISM, a skeleton-guided conditional latent diffusion framework reconstructing dense human point clouds from sparse mmWave radar alone. SkelKD regresses radar-aligned joint positions via learnable cross-attention. GeoAE produces density-invariant coarse body geometry. Together with action class embeddings, these three conditioning streams guide a Dynamic Transformer denoiser under a VP-SDE framework. On MIST, PRISM achieves COV-CD 0.362 against 0.113 for the strongest completion baseline; under through-obstacle attenuation where all baselines collapse to fixed-template outputs, PRISM maintains meaningful reconstruction.

Method

Two-path skeleton-guided architecture

The condition path processes the sparse mmWave cloud through SkelKD, GeoAE, and an action class embedding, assembling three streams into a 145-token K/V sequence injected at every denoising layer. The diffusion path encodes the cloud into a compact 8×120 structured latent via HierVAE, iteratively denoises it under a VP-SDE schedule driven by the condition K/V sequence through the Dynamic Transformer, and decodes the result to a dense 256-point output cloud.

Figure 1. Overview of PRISM's two-path architecture for dense human point cloud reconstruction from sparse mmWave radar.

SkelKD

Takes sparse mmWave cloud as sole input and regresses 17 radar-aligned skeleton joints via 17 learnable query embeddings attending over per-point features. Trained with MPJPE loss and noise/masking augmentation to simulate occlusion; frozen before diffusion training.

HierVAE

Hierarchical VAE that encodes sparse point clouds into a compact 8×120 structured latent Z₀ through six levels of local grouping and residual self-attention. Decoder produces K=256 output points. Pre-trained and frozen so the denoiser trains within a stable latent geometry.

GeoAE + Condition Network

GeoAE learns a density-invariant mapping from sparse radar cloud to 128 canonical coarse body points, stable under occlusion. The Condition Network encodes GeoAE output and SkelKD GCN features into a 145-token K/V sequence; action identity enters the global AdaLN signal c = e_t + e_y + Linear(g_s).

Dynamic Transformer

Six ResidualBlocks alternating cross-attention (blocks 1, 3, 5) with self-attention (blocks 2, 4, 6), conditioned on the 145-token K/V sequence. AdaLN driven jointly by timestep, action label, and skeleton global pool. 200-step CFG inference yields X_f ∈ ℝ^(256×3).

Results

Quantitative & qualitative evaluation

Evaluated on MIST-OL0 (test subjects P07–P08, all 30 action classes) using distribution metrics: COV-CD ↑ (fraction of ground-truth shapes covered) and MMD-CD ↓ (fidelity; ×10⁻³). All baselines re-trained on MIST subjects P01–P06.

Method	Venue	Type	COV-CD ↑	MMD-CD ↓ (×10⁻³)
PCN	3DV'18	Completion	0.083	9.01
PoinTr	ICCV'21	Completion	0.113	8.26
SnowflakeNet	ICCV'21	Completion	0.079	10.72
LAKe-Net	CVPR'22	Completion	0.085	8.69
mmPoint	BMVC'23	Radar	0.097	8.91
LION	NeurIPS'22	Generative	0.185	7.37
TIGER	CVPR'24	Generative	0.209	7.08
PRISM (Ours)	—	Generative	0.362	4.91

PRISM achieves 3.2× coverage improvement over PoinTr and 40.6% lower matching distance.

Dataset

Multi-modal human sensing datasets

MIST Coming Soon

First mmWave human point cloud dataset with a physical barrier between sensor and subject. Paired clear-view (OL0) and through-obstacle (OL1, OL2) captures with synchronized overhead LiDAR as dense 3D ground truth. Dual TI IWR6843 radars, Livox Mid-360, and Intel RealSense D435i.

8 subjects · 30 action classes · 3 occlusion levels · ~640k paired frames. To be released upon paper acceptance.

MM-Fi Public

Large-scale multi-modal dataset with synchronized mmWave, LiDAR, RGB-D, and WiFi CSI. 40 subjects · 27 actions · 4 environments · 320K+ frames.

Cross-dataset transfer: E01 · S08–S10 · 8 shared action classes (A03, A12, A13, A17, A19, A22, A26, A27).

Citation

Cite this work

If you find PRISM useful in your research, please consider citing:

@inproceedings{prism2026,
  title     = {PRISM: Human Point Cloud Reconstruction via
               Skeleton-Guided Diffusion from MmWave Radar},
  author    = {Anonymous},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

PRISM: Human Point Cloud Reconstruction via Skeleton-Guided Diffusion from mmWave Radar

From sparse radar signalsto complete human shapes