Title: HyperDet: Advancing 3D Object Detection via Hyper 4D Radar Point Clouds

Abstract:

What is the ultimate capability of 3D object detection when relying exclusively on 4D radar? Although 4D radar provides autonomous perception systems with velocity-sensitive data and resilience against adverse weather conditions, its practical application is hindered by point clouds that are often sparse, noisy, and temporally unstable. To address these limitations, we introduce HyperDet, a detector-agnostic framework designed to generate task-specific "hyper" 4D radar point clouds prior to the detection phase.

HyperDet begins by enhancing short-window, surround-view radar observations through a process that includes spatio-temporal accumulation, validation across sensors, and motion compensation guided by Doppler data. This initial step significantly boosts the reliability of signal returns and ensures temporal consistency. Subsequently, the framework employs a foreground generative enhancement technique. By leveraging LiDAR-guided pseudo-radar supervision—which is accessible only during the training phase—HyperDet enriches the geometric details of objects while maintaining the integrity of the measured radar background and native radar characteristics.

To ensure robustness during the training of detectors, HyperDet incorporates radar-aware object-level augmentation, which preserves Doppler consistency even when objects are geometrically relocated. At the inference stage, the system operates using radar data exclusively and can be seamlessly integrated with conventional 3D detectors. Our experiments, conducted on two public surround-view 4D radar datasets, show that HyperDet delivers consistent performance gains over raw radar inputs across various standard 3D detectors. These results confirm that enhancing radar inputs at the data level is a highly effective strategy for improving radar-only 3D detection.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC