基于置信域伪标签策略的半监督三维目标检测

Trust region pseudo-labels strategy for semi-supervised 3D object detection

当前基于点云的三维目标检测方法很大程度上依赖于大规模高质量的三维标注。为了减少所需标签量,基于SESS网络提出了一种新的三维目标检测方法:基于置信域伪标签策略的半监督三维目标检测。首先设计了一种置信域伪标签策略,将学生网络的输出分成有标签和无标签两部分,有标签部分利用ground truth进行全监督学习,无标签部分基于教师网络的类别和对象预测置信分数,利用一种有效的过滤机制,筛选出高质量教师预测,并转换成相应伪标签,用于监督学生网络无标签部分。其次,设计了一个TransVote模块,通过Transformer机制,增强每个点云与其邻域点之间的相互注意,聚合点云局部特征。在10%标记数据、mAP@0.25下,该算法在ScanNetV2和SUN RGB-D数据集上分别超越了基准线8.63%、6.75%,显著提高了半监督三维目标检测算法的检测精度。

Current point cloud-based 3D object detection methods largely rely on large-scale,high-quality 3D annotations.In order to reduce the required amount of labels,this paper proposed a new 3D object detection method based on the SESS network,trust region pseudo-supervised strategy for semi-supervised 3D object detection.Firstly,this paper designed a trust region pseudo-supervised strategy,which divided the output of the student network into labeled and unlabeled parts.The labeled part used ground truth for supervised learning,and based on the class and object prediction confidence score of the teacher network,the unlabeled part used an effective filtering mechanism to filter out high-quality teacher predictions and converted them into corresponding pseudo-labels for supervising the unlabeled part of the student network.Secondly,this paper designed a TransVote module,which used the Transformer mechanism to enhance the mutual attention between each point cloud and its neighboring points and aggregated the local features of the point cloud.With 10%labeled data and mAP@0.25,this method surpasses the baseline by 8.63%and 6.75%on the ScanNetV2 and SUN RGB-D datasets,which significantly improves the detection accuracy of semi-supervised 3D object detection algorithm.