GeoTeacher: Geometry-Guided Semi-Supervised 3D Object Detection

TL;DR

GeoTeacher addresses the limited geometric sensitivity in semi-supervised 3D object detection by introducing geometric relation supervision and distance-aware voxel-wise augmentation. It transfers object geometry knowledge from a teacher to a student via keypoint-based geometric relations and enhances geometric diversity through voxel-level augmentations with a distance-decay mechanism. The approach is compatible with existing SS3D methods and demonstrates state-of-the-art-like gains on ONCE and Waymo across varying labeling regimes. The findings highlight the value of incorporating higher-order geometric structure into SSL for 3D perception with limited labels.

Abstract

Semi-supervised 3D object detection, aiming 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 consistency 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 information, 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 augmentation strategy that increases the diversity of object geometries, 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. Code will be available at https://github.com/SII-Whaleice/GeoTeacher

Figures (2)

• Figure 1: The framework of our GeoTeacher. After applying strong augmentation, we further introduce our proposed data augmentation strategy to enhance the geometric diversity of objects. The student model is supervised with both the standard semi-supervised loss and our geometric relation supervision loss, which effectively improves the student model’s ability to comprehend the intrinsic geometric information of objects.
• Figure 2: Some visualization examples from the ONCE dataset NEURIPS67c6a1e7 are presented. From top to down, each row displays the baseline results and the predictions of our GeoTeacher. Red rectangles denote ground truth, black rectangles represent predictions, and red dashed circles highlight the improvements achieved by our method.