Class-Aware PillarMix: Can Mixed Sample Data Augmentation Enhance 3D Object Detection with Radar Point Clouds?

TL;DR

This paper tackles data scarcity in radar-based 3D object detection by evaluating mixed sample data augmentation (MSDA) methods originally designed for LiDAR, identifying radar-specific challenges, and proposing Class-Aware PillarMix (CAPMix). CAPMix performs pillar-wise MixUp with per-pillar, class-aware mix ratios drawn from Beta distributions, selectively applying these ratios to pillars containing labeled objects to preserve essential structure while increasing sample diversity. Across Bosch Street and K-Radar datasets, CAPMix outperforms LiDAR-oriented MSDA approaches and achieves notable data-efficiency, with 25% of the data using CAPMix matching or exceeding the performance of 50% data without MSDA, and 50% CAPMix surpassing the full-data baseline in several metrics. These results demonstrate CAPMix’s effectiveness in radar data augmentation and its potential to reduce data collection costs, while motivating further exploration of parameter learning and extensions to semi-supervised and domain adaptation settings.

Abstract

Due to the significant effort required for data collection and annotation in 3D perception tasks, mixed sample data augmentation (MSDA) has been widely studied to generate diverse training samples by mixing existing data. Recently, many MSDA techniques have been developed for point clouds, but they mainly target LiDAR data, leaving their application to radar point clouds largely unexplored. In this paper, we examine the feasibility of applying existing MSDA methods to radar point clouds and identify several challenges in adapting these techniques. These obstacles stem from the radar's irregular angular distribution, deviations from a single-sensor polar layout in multi-radar setups, and point sparsity. To address these issues, we propose Class-Aware PillarMix (CAPMix), a novel MSDA approach that applies MixUp at the pillar level in 3D point clouds, guided by class labels. Unlike methods that rely a single mix ratio to the entire sample, CAPMix assigns an independent ratio to each pillar, boosting sample diversity. To account for the density of different classes, we use class-specific distributions: for dense objects (e.g., large vehicles), we skew ratios to favor points from another sample, while for sparse objects (e.g., pedestrians), we sample more points from the original. This class-aware mixing retains critical details and enriches each sample with new information, ultimately generating more diverse training data. Experimental results demonstrate that our method not only significantly boosts performance but also outperforms existing MSDA approaches across two datasets (Bosch Street and K-Radar). We believe that this straightforward yet effective approach will spark further investigation into MSDA techniques for radar data.

Figures (6)

• Figure 1: Comparison between CAPMix and no MSDA reveals remarkable improvements on the Bosch Street dataset.
• Figure 2: (a) Rotating points disrupts the radiation pattern. (b) Points from different radar sensors adhere to their respective local polar distributions. (c) Mixing points based on elevation angles may be ineffective due to the low angular resolution of radar sensors.
• Figure 3: Examples of LiDAR-based MSDA methods applied to radar point clouds in Bosch Street dataset. The point clouds are aggregated from five sensors. In PolarMix, a 90$^\circ$ region along the azimuth axis is extracted and swapped between $s^A$ and $s^B$, while object-level points from $s^B$ are copied and rotated into the new sample $\tilde{s}$. In LaserMix, the point clouds are divided into six regions, and points from non-adjacent areas are combined from $s^A$ and $s^B$.
• Figure 4: Overall framework comparison of PillarMix and our proposed CAPMix. Both methods share the pillarization step, but PillarMix simply switches pillars between two samples $s^A, s^B$. In contrast, CAPMix randomly selects pillars and then performs class-aware MixUp using different $\beta$ distributions $\textcolor{sparse}{\beta_{s}},\textcolor{moderate}{\beta_{m}},\textcolor{dense}{\beta_{d}}$ based on the classes in the pillars.
• Figure 5: Comparison of the mixing effect, where grid cells represent pillars and numbers indicate the mix ratio $\lambda$.