GeoTeacher: Geometry-Guided Semi-Supervised 3D Object Detection
Jingyu Li, Xiaolong Zhao, Zhe Liu, Wenxiao Wu, Li Zhang
AI summary
Problem
Existing semi-supervised 3D object detection methods struggle to capture crucial object geometric structures when labeled data is scarce, limiting detection accuracy and localization. Prior approaches focus mainly on pseudo-label quality or feature consistency, overlooking higher-order geometric relations.
Approach
GeoTeacher transfers geometric knowledge from a teacher to a student model via a keypoint-based relation supervision module, while a distance-decay voxel-wise augmentation strategy diversifies object geometries during training. This plug-and-play framework explicitly guides the student to learn intrinsic object structures from unlabeled data.
Key results
- Keypoint-based geometric relation supervision module
- Distance-decay voxel-wise data augmentation strategy
- Consistent mAP gains of +1.76 to +3.02 over baselines
- New state-of-the-art results on ONCE and Waymo datasets
Why it matters
Enables robust 3D object detection with minimal labeled data, directly advancing autonomous driving and robotic perception systems.
Abstract
Semi-supervised 3D object detection (SS3D), aim- ing to explore unlabeled data for boosting 3D object detectors, has emerged as an active research area in recent years. Some previous methods have shown substantial improvements by either employing heterogeneous teacher models to provide high- quality pseudo labels or enforcing feature-perspective consis- tency between the teacher and student networks. However, these methods overlook the fact that the model usually tends to exhibit low sensitivity to object geometries with limited labeled data, making it difficult to capture geometric infor- mation, which is crucial for enhancing the student model’s ability in object perception and localization. In this paper, we propose GeoTeacher to enhance the student model’s ability to capture geometric relations of objects with limited training data, especially unlabeled data. We design a keypoint-based geometric relation supervision module that transfers the teacher model’s knowledge of object geometry to the student, thereby improving the student’s capability in understanding geometric relations. Furthermore, we introduce a voxel-wise data aug- mentation strategy that increases the diversity of object geome- tries, thereby further improving the student model’s ability to comprehend geometric structures. To preserve the integrity of distant objects during augmentation, we incorporate a distance- decay mechanism into this strategy. Moreover, GeoTeacher can be combined with different SS3D methods to further improve their performance. Extensive experiments on the ONCE and Waymo datasets indicate the effectiveness and generalization of our method and we achieve the new state-of-the-art results.