| Deep learning has been successfully applied to solve various computational vision problems.But in the application scenarios of autonomous driving,robotics,and augmented reality,images cannot provide all the information needed to perceive the environment.Compared with images,3D data can restore real scenes more accurately.Under the impetus of different applications,3D data processing has gradually become a hot research area.Among them,point cloud is a relatively concise representation of 3D data,whether in acquisition or application.Therefore,analyzing the point cloud efficiently and accurately is the key to let the machine understand 3D data.However,the existing methods have many problems.For example,in the point cloud scene semantic segmentation task,many methods lack sufficient context awareness and cannot correctly segment different objects with similar structures,such as door and window.In addition,the success of the current method mainly relies on a large amount of labeled data.However,in practice,the cost of point cloud labeling is extremely expensive.At the same time,in point cloud classification task,due to the heavy reliance on a large amount of training data,the generalization ability of existing models is insufficient,and it is difficult to generalize to new categories.This thesis focuses on the above three issues,and the research content is as follows:First,for existing methods that are easy to confuse objects with the same structure but different categories,a large-scale point cloud semantic segmentation model based on multi-level context is designed.The idea of self-calibration is adopted to explicitly model the context between different scales.At the same time,we propose a semanticrelated context module to constrain the features of points of the same category.The experimental results in multiple datasets show that the proposed method can effectively avoid the ambiguity caused by objects with similar structures,and the performance is superior to the state-of-the-art methods.Second,in view of the expensive labeling cost,in this thesis,we discuss the labeling methods used by the existing weakly supervised point cloud segmentation methods,and propose a simpler labeling method.Meanwhile,aiming at the problem of insufficient supervision,a novelty model for weakly supervised point cloud segmentation is designed.We take point cloud completion as an auxiliary task to strengthen the model’s ability to perceive the relation between part and the whole of the sample without introducing additional labels.In addition,in order to make full use of labels,a category attention module is proposed in thesis,which utilizing category prototypes to further improve the segmentation results.Experiments show that our method has superior performance compared with current methods and is extremely competitive on several datasets.Third,aiming at the problem of insufficient generalization of existing models in point cloud classification,a point cloud classification model based on few-shot learning is proposed.In this thesis,metric learning is used to solve the problem of few-shot learning.In order to improve the feature extraction capabilities,the self-supervised framework BYOL is used for pre-training,and the learned knowledge is transferred to the point cloud classification task.Additionally,a global contextual reasoning module is proposed to improve results.Specifically,it help model to make a more reasonable judgement by taking the feature of each category into account.The experimental results show that the proposed method is highly robust and generalizable,and can efficiently solve the problem of point cloud classification under few-shot learning condition. |
