Semi-supervised 3D Target Detection Method Based On Comprehensive Confidence Threshold

3D object detection is an important task in the field of computer vision. Most of the current methods require a large amount of high-quality 3D annotation data. Due to the high cost of annotation, complex scenes, sparse point clouds and other reasons, the research becomes more difficult. Although semi supervised learning can use a large amount of unlabeled data and a small amount of labeled data to improve model learning, the method of pseudo label screening in semi supervised learning still needs to be improved. Therefore, we propose a confidence threshold-based screening method for pseudo-labeling for indoor 3D target detection tasks. We design a method to perform multiplicative fusion of multiple confidence thresholds to improve the quality of screened pseudo-labels. In addition, we design a neighborhood relationship between some of the class groups those are easy to be misdetected, so as to improve the detection accuracy. The effectiveness of our method is demonstrated on the SUNRGB-D dataset.