Auto-Labelling-Based Domain Transfer for 3D Object Detection on a Bicycle-Mounted LiDAR Platform

As autonomous driving research expands to include bicycle-mounted sensor platforms, a significant challenge arises: detectors trained on automotive data struggle to generalize to the unique perspective, sensor mounting, and point density of a bicycle. This paper investigates whether these vehicle-trained detectors can be effectively adapted to the cyclist domain using only automatically generated labels, bypassing the need for expensive manual annotation.

The authors introduce the FUSE-Bike benchmark, which consists of 1,027 LiDAR keyframes recorded in urban Munich. They utilize a specialized auto-labeling pipeline (VRU-Label3D) to generate training data for three classes: vehicles, pedestrians, and cyclists. To evaluate performance, they compare four diverse, nuScenes-pre-trained 3D object detectors (CenterPoint, SECOND-MH, TransFusion-L, and VoxelNeXt) in two settings: zero-shot (original weights) and finetuned (adapted on auto-labels). The evaluation is conducted against a held-out set of 86 manually verified ground-truth keyframes.

The study reveals a substantial domain gap, particularly for vulnerable road users (VRUs) like pedestrians and cyclists, which are under-represented in standard automotive datasets. Finetuning on auto-labels consistently improves performance across all tested architectures, with mAP gains ranging from 13.7 to 23.4 points. Notably, the adapted detectors do not just mimic the auto-labels; they generalize beyond them, achieving higher precision and recall than the labels used during training. This demonstrates that auto-labeling is a viable, scalable substitute for manual annotation when adapting perception systems to new, unconventional platforms.

This work provides a reproducible baseline for VRU-centric 3D detection and offers a practical path for researchers to deploy high-performance perception on mobile platforms without the prohibitive cost of manual 3D labeling. By showing that detectors can improve upon their own training signals, the paper highlights the potential for self-improving perception systems in complex urban environments.