GBlobs: Local LiDAR Geometry for Improved Sensor Placement Generalization

TL;DR

GBlobs address the generalization gap in LiDAR-based 3D detection caused by reliance on absolute coordinates by encoding local neighborhoods as Gaussian blobs with mean $\mu = (1/N) \sum_i p_i$ and covariance $\Sigma = (1/N) \sum_i (p_i - \mu)(p_i - \mu)^T$. To handle far-range sparsity, a dual-model system runs a GBlobs-based detector alongside a global Cartesian detector, with Test-Time Augmentation and NMS, and predictions are fused using a distance threshold of $\delta_d = 30$ m so that GBlobs dominate near-range detections while global predictions handle farther objects. Experiments on RoboSense 2025 Track 3 with synthetic CARLA data show that replacing global inputs with GBlobs yields significant gains ($+1.67$ AP) that are further boosted to $+2.79$ AP with TTA, culminating in top placement in the challenge. This work demonstrates the potential of local geometric features to enhance generalization across sensor placements, offering a practical path toward robust, placement-agnostic 3D detection systems for autonomous sensing.

Abstract

This technical report outlines the top-ranking solution for RoboSense 2025: Track 3, achieving state-of-the-art performance on 3D object detection under various sensor placements. Our submission utilizes GBlobs, a local point cloud feature descriptor specifically designed to enhance model generalization across diverse LiDAR configurations. Current LiDAR-based 3D detectors often suffer from a \enquote{geometric shortcut} when trained on conventional global features (\ie, absolute Cartesian coordinates). This introduces a position bias that causes models to primarily rely on absolute object position rather than distinguishing shape and appearance characteristics. Although effective for in-domain data, this shortcut severely limits generalization when encountering different point distributions, such as those resulting from varying sensor placements. By using GBlobs as network input features, we effectively circumvent this geometric shortcut, compelling the network to learn robust, object-centric representations. This approach significantly enhances the model's ability to generalize, resulting in the exceptional performance demonstrated in this challenge.

Figures (3)

• Figure 1: Our method first takes an input point cloud ① and generates a batch of randomly augmented frames ②. This batch is then inferred ③ by two models: the BEVFusion-L liu2023bevfusion baseline model and the same model trained with GBlobs cvpr2025gblobs. The augmented predictions are then reversed (de-augmented) and combined using Non-Maximum Suppression (NMS) ④. Finally, the predictions from the two models are fused ⑤ via $\delta_t$ fusion, where the GBlobs-trained predictions are used up to the $\delta_t$ distance threshold, and the standard Cartesian model predictions are used beyond it. Best viewed on a monitor and zoomed in for detail.
• Figure 2: Exemplary LiDAR frames from the RoboSense Challenge 2025, Track 3: Sensor Placement dataset.
• Figure 3: We evaluate two BEVFusion-L liu2023bevfusion models: a standard model trained with global input features (\ref{['fig:tikz:global']}) and a model trained with GBlobs cvpr2025gblobs (\ref{['fig:tikz:gblobs']}). The $x$-axis represents the distance cut-off, where predictions and ground truth below this value are ignored during evaluation. The $y$-axis reports the mean Average Precision (mAP) computed across all detection classes on the validation split. Both models were trained on the training split of the RoboSense 2025 Challenge: Track 3 dataset.