Understanding the Domain Gap in LiDAR Object Detection Networks

Richter, Jasmine; Faion, Florian; Feng, Di; Becker, Paul Benedikt; Sielecki, Piotr; Glaeser, Claudius · 2022 · OpenAlex-citations

DOI: 10.48550/arxiv.2204.10024

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper investigates the domain gaps between high-resolution and low-resolution LiDAR sensors in object detection networks for autonomous driving. The motivation stems from the impracticality of collecting and annotating training datasets that fully represent the diverse, changing open-world for every new sensor generation. To address this, the authors aim to understand sensor-to-sensor domain gaps to facilitate the reuse and improvement of datasets and models. The study uniquely isolates these gaps from other variables like weather or location by using a dataset where high-resolution (Hesai Pandar64P) and low-resolution (Velodyne VLP-32C) sensors recorded data simultaneously. The experimental design utilized a dataset collected over 7,800 km across 33 cities in six European countries, comprising approximately 13,100 frames and 432,000 labeled bounding boxes. The sensors were mounted in an alternating pattern to ensure full 360-degree coverage. The authors employed an object detection network based on the PIXOR architecture, using identical random initializations and hyperparameters for all trials. They quantified performance using relative average precision (AP) and analyzed the relationship between recall and the average number of LiDAR points received per object. The study distinguishes between two specific gaps: the inference domain gap (training on one sensor, evaluating on both) and the training domain gap (training on different sensors, evaluating on one). The results reveal two distinct behaviors. For the inference domain gap, performance strongly depends on the number of points per object at inference time. Using a higher-resolution LiDAR during inference improved detection performance regardless of the training data source, as higher resolution provides more points per object, increasing recall. Conversely, the training domain gap showed no such dependence on point density. Training with higher-resolution data did not improve performance; instead, networks performed best when trained on data from the same sensor type used for evaluation. Specifically, training on high-resolution data while evaluating on low-resolution data yielded worse results than training and evaluating on low-resolution data. Spatial analysis confirmed that inference performance varied with the spatial distribution of point density, whereas training performance did not. The significance of these findings lies in the conclusion that distinct strategies are required to close inference and training domain gaps. For inference, upgrading to higher-resolution sensors can mitigate domain gaps by increasing point density. However, for training, simply using higher-resolution data is insufficient and potentially detrimental if the deployment sensor is lower-resolution. The authors suggest that domain adaptation techniques must account for these differences, noting that similar behaviors were observed in preliminary weather domain gap experiments. This work provides a foundational understanding of how sensor resolution impacts neural network generalization in LiDAR-based perception systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness