Instance-Level Strong Augmentation for Semi-Supervised 3D Object Detection / under review at CVPR 2025

Semi-supervised 3D object detection aims to improve model performance by leveraging both labeled and unlabeled data. Existing methods primarily focus on scene-level augmentations, such as rotation, flipping, and scaling, to enhance the training of student models. However, scene-level augmentations fail to fully exploit instance-specific information, which is essential for accurate object detection in 3D environments. In this paper, we propose ISA, Instance-level Strong Augmentation strategy, for semi-supervised 3D object detection. ISA includes three key augmentation strategies: instance switch, intra-class mixup, and inter-class mixup. These strategies enable the model to better leverage instance-specific features, improving the learning performance over unlabeled data. To ensure consistent and reliable learning, we also introduce augmentation constraints, including instance box fitting and density-controlled instance generation. These innovations work together to enhance the model’s ability to generalize across diverse scenarios. Extensive experiments on the ScanNet and SUN RGB-D datasets show that our method consistently outperforms baseline models, achieving significant improvements in detection accuracy and generalization, particularly in low-labeled data settings.