PART

Overview

Detect motion beyond the training taxonomy.

Closed-set detectors can overlook rare moving objects simply because they were never assigned a familiar category. PART uses radar's direct Doppler measurements as category-independent physical evidence: it asks whether an object is moving, not whether it looks like a class seen during training.

Operating directly on temporally aggregated sparse radar points, PART predicts an existence confidence, representative surface point, and ground-plane velocity for every moving-object hypothesis.

Main results on nuScenes

Compact model, motion-first output.

1.1M

Parameters

0.8827

Class-agnostic AP

0.3188 m

Surface translation error

0.8084 m/s

Velocity error

0.9203

Recall on rare objects

Method

Physics guides every association.

01

Doppler-Aware Query Initialization

Forms input-dependent object queries by grouping radar returns that agree in position and velocity.

02

Physics-Guided Cross-Attention

Uses radial-Doppler consistency and radar cross section to associate queries with meaningful returns.

03

Uncertainty-Aware Supervision

Avoids treating plausible radar-supported hypotheses as hard negatives when annotations are incomplete.

In the wild

Motion evidence persists where appearance does not.

Adverse illuminationNight-time driving scene with headlight glare.

Text description of this demonstration

A forward-driving sequence is shown at night, where sparse street lighting and headlight glare make visual appearance unreliable. Across the frames, teal markers denote moving radar returns. Blue surface markers and arrows show PART's predicted moving-object location and ground-plane velocity. The sequence illustrates that radar motion evidence remains available as the illuminated vehicle moves through the dark and glare-affected scene.

Rare moving objectPowered wheelchair detection beyond the usual taxonomy.

Text description of this demonstration

A powered wheelchair user moves through a road scene. This category is uncommon in closed-set driving taxonomies, yet the radar observations form a coherent moving pattern. Teal markers identify moving radar returns, while blue surface markers and arrows indicate PART's predicted surface location and ground-plane velocity. The demonstration highlights class-agnostic detection driven by physical motion rather than a familiar object label.

Qualitative results

From rare objects to incomplete labels.

special-category objects moving ground truth PART prediction

Powered wheelchair user

Personal-mobility rider behind a guardrail detected by PART. — Personal-mobility rider

Dog-walking pedestrian at night detected by PART. — Pedestrian at night

Severely occluded crossing vehicle detected by PART. — Occluded crossing vehicle

Severely occluded moving vehicle detected by PART. — Severely occluded vehicle

Rainy night vehicle under headlight glare detected by PART. — Rain and headlight glare

Occluded crossing vehicle at night detected by PART. — Night-time occlusion

Unannotated vehicle in rain detected by PART. — Unannotated vehicle in rain

Unannotated vehicle under headlight glare detected by PART. — Unannotated under glare

Unannotated oncoming vehicle in rain detected by PART. — Unannotated oncoming vehicle

Unannotated motorcycle under headlight glare detected by PART. — Unannotated motorcycle

Quantitative results

Compact, robust, and class-agnostic.

All values below are reproduced from the current manuscript. Scroll horizontally to inspect wide benchmark tables.

Table 1Overall comparison with existing 3D detectors.

Method	Params.	Mod.	CA-AP ↑	mASTE ↓	mAVE ↓	mATE ↓
CenterPoint	23M	L	0.8746	0.4988	0.7097	0.4542
VoxelNeXT	31M	L	0.8801	0.4498	0.6166	0.4172
TransFusion-L	32M	L	0.8865	0.3617	0.8600	0.2790
FCOS3D	55M	C	0.5254	1.2274	2.6908	1.2818
PGD	56M	C	0.5423	1.4796	2.7007	1.3143
PETR	83M	C	0.6553	2.9833	2.3070	0.9340
BEVFusion	157M	LC	0.8877	0.3594	0.8138	0.2772
LiRaFusion	23M	LR	0.8865	0.5282	0.7556	0.4792
CRTFusion	81M	RC	0.8612	0.6699	0.7711	0.6206
DAQI Proposals	N/A	R	0.6776	0.5884	1.5758	1.7806
PillarNet-R	16M	R	0.7492	0.9586	0.8702	0.8919
RadarDistill	41M	R(L)	0.7684	0.7764	0.7499	0.7587
PART	1.1M	R	0.8827	0.3188	0.8084	1.4846

Best value in bold; next two best distinct values are underlined. L: lidar; C: camera; R: radar. DAQI Proposals use DBSCAN cluster centroids and mean velocities directly. RadarDistill uses a lidar teacher during training but radar input at inference.

Table 2Robustness comparison under challenging conditions.

Method	Mod.	Night			Rainy			Night + Rainy
Method	Mod.	CA-AP ↑	mASTE ↓	mAVE ↓	CA-AP ↑	mASTE ↓	mAVE ↓	CA-AP ↑	mASTE ↓	mAVE ↓
CenterPoint	L	0.8297	0.3398	0.7795	0.8528	0.5671	0.6729	0.8107	0.3410	0.7752
VoxelNeXT	L	0.8361	0.3602	0.6968	0.8606	0.4503	0.5776	0.7861	0.4930	0.6652
TransFusion-L	L	0.8584	0.2530	0.6825	0.8662	0.3699	0.8649	0.8520	0.3122	0.6785
FCOS3D	C	0.5197	0.6753	2.0607	0.5203	1.3734	2.6823	0.4973	0.6429	2.3148
PGD	C	0.6035	0.6951	2.6314	0.5114	1.2124	2.5529	0.5704	0.5634	1.9414
PETR	C	0.6855	1.6695	2.3924	0.6978	3.2089	2.0686	0.7257	1.4301	2.5327
BEVFusion	LC	0.8665	0.2533	0.7634	0.8754	0.4137	0.8732	0.8768	0.3278	0.7924
LiRaFusion	LR	0.8816	0.3382	0.7499	0.8864	0.5756	0.8066	0.8793	0.3007	0.5332
CRTFusion	CR	0.8683	0.4730	0.8821	0.8480	0.7055	1.0037	0.8863	0.3864	1.0296
DAQI Proposals	R	0.7026	0.4104	1.2325	0.6789	0.7148	1.6498	0.6889	0.4865	1.2728
PillarNet-R	R	0.8084	0.6602	0.9246	0.7711	1.0402	1.0874	0.8266	0.6334	0.6675
RadarDistill	R(L)	0.8162	0.3763	0.8175	0.7996	0.8528	0.7877	0.8164	0.4360	2.0693
PART	R	0.8930	0.2585	0.4651	0.8916	0.2811	0.7590	0.8892	0.2034	0.2994

Table 3Comparison on rare and safety-relevant object categories.

Method	Mod.	Police Vehicles N_GT = 70			Wheelchair N_GT = 260			Personal Mobility N_GT = 102			Animal N_GT = 13
Method	Mod.	Recall ↑	ASTE ↓	AVE ↓	Recall ↑	ASTE ↓	AVE ↓	Recall ↑	ASTE ↓	AVE ↓	Recall ↑	ASTE ↓	AVE ↓
CenterPoint	L	0.8286	0.2633	0.5565	0.0346	0.4032	0.3574	0.1863	0.2331	0.4617	0.1538	0.0979	0.2048
VoxelNeXT	L	0.8571	0.2872	0.6912	0.0115	0.3614	0.4953	0.1765	0.1923	0.3328	0.0000	N/A	N/A
TransFusion-L	L	0.5857	0.2038	0.6645	0.0038	0.4830	0.1431	0.0196	0.1621	0.4392	0.0000	N/A	N/A
FCOS3D	C	0.5286	0.5357	3.9904	0.0000	N/A	N/A	0.0588	0.2053	3.3621	0.0000	N/A	N/A
PGD	C	0.6143	0.4390	3.8758	0.0038	0.2099	1.5279	0.0392	0.1553	3.1449	0.0000	N/A	N/A
PETR	C	0.0000	N/A	N/A	0.0000	N/A	N/A	0.0000	N/A	N/A	0.0000	N/A	N/A
BEVFusion	LC	0.6000	0.3038	0.6558	0.0000	N/A	N/A	0.0000	N/A	N/A	0.0000	N/A	N/A
LiRaFusion	LR	0.5143	0.2723	0.7073	0.0077	0.1715	0.2147	0.0000	N/A	N/A	0.0000	N/A	N/A
CRTFusion	CR	0.6286	0.5582	0.7085	0.0192	0.2666	0.4843	0.1471	0.1546	1.6308	0.0000	N/A	N/A
DAQI Proposals	R	0.8571	0.5975	2.3529	0.8808	0.3043	0.3735	0.4706	0.1331	1.3884	0.6154	0.1305	0.1662
PillarNet-R	R	0.5588	0.4506	0.9424	0.0115	0.1633	0.2811	0.0000	N/A	N/A	0.0000	N/A	N/A
RadarDistill	R(L)	0.7647	0.3795	0.7877	0.0154	0.2340	0.4015	0.0000	N/A	N/A	0.0000	N/A	N/A
PART	R	0.9571	0.2931	1.0003	0.9577	0.1142	0.2161	0.8039	0.1063	0.6048	0.9231	0.2266	0.1838

N_GT is the number of eligible ground-truth boxes. Gray cells indicate zero recall despite ground-truth instances; N/A denotes an undefined error because no true positive is available.

Table 4Component ablation.

Exp.	CA-AP ↑	mASTE ↓	mAVE ↓	Special Recall ↑
w/o DAQI	0.6198	0.6777	1.3048	0.6646
w/o PGCA	0.8815	0.3300	0.8348	0.8931
w/o UAS	0.7948	0.3149	0.8292	0.3899
w/o PWSH	0.8827	0.3607	0.8325	0.9182
FULL	0.8827	0.3188	0.8084	0.9203

The w/o DAQI variant is reported at a fixed score threshold of 0.2 because it produces no valid detections at the standard 0.5 threshold; CA-AP uses the full confidence sweep. UAS: Uncertainty-Aware Supervision. PWSH: Point-Wise Surface Head. Special recall is evaluated over rare and safety-relevant categories.

Citation

If this work is useful to you, please consider citing our paper.

@article{sun2026part,
  title   = {If It Moves, Radar Knows: A Physics-Aware Radar Transformer for Class-Agnostic Moving-Object Detection},
  author  = {Sun, Yinghao and Li, Shuguang and Shao, Jinliang and Li, Tieshan},
  journal = {arXiv preprint arXiv:2609.02289},
  year    = {2026}
}