| Recent advances in machine learning and related computational technologies have led to significant progress in deep learning-based algorithms for processing 3D point clouds.These algorithms have found growing utility in diverse fields,including but not limited to autonomous driving,robot vision,and intelligent manufacturing.Although deep-learning-based artificial neural networks have demonstrated remarkable capabilities in learning multi-level abstract feature representations for a variety of tasks such as 3D point cloud registration,classification,and object detection,enhancing their performance usually necessitates leveraging data-driven approaches.Regrettably,the acquisition of point cloud data,particularly manual annotation,can incur higher costs relative to other domains such as natural language processing and image processing,thereby posing significant obstacles to some envisaged applications.This thesis aims to reduce the dependence of deep point cloud processing algorithms on labeled data and thoroughly investigate the issue through self-supervised representation learning and transfer learning of point clouds.The first part of this thesis proposes a self-supervised representation learning method that utilizes the Gaussian Mixture Model(GMM)theory in the point cloud registration task.The proposed method considers two frames of point clouds as samples from a unified GMM.Then the point cloud registration task is modeled as a problem of reconstructing the unified GMM,and solving the rigid transformation simultaneously.Additionally,a novel feature interaction extraction module is proposed to fully integrate information from the two point clouds for reconstructing a higher-quality GMM.Experiments demonstrate that the models obtained using this self-supervised method can achieve performance comparable to or even surpass the current state-of-the-art point cloud registration method based on supervised learning.After obtaining a pre-trained model through representation learning,it is essential to efficiently apply it to downstream tasks.Inspired by research into the application of the Transformer architecture to both point clouds and images,this thesis investigates the transfer of the CLIP Transformer to downstream tasks of point clouds.Specifically,the proposed method uniformly divides point clouds into discrete point patches using farthest point sampling and k-nearest neighbor algorithms,then extracts features for each patch using a point cloud tokenizer module to ensure that both the point clouds and images have identical representation forms in the feature space.The extracted features are concatenated with a few learnable feature vectors and fed into the frozen CLIP Transformer model for updating,thereby progressively aligning the point clouds with semantically similar images in the feature space.Finally,the updated features are input into different task decoding modules to obtain predictions for each task.Extensive experiments demonstrate that when compared to other similar approaches for pre-training on point cloud data,this method can deliver performance equivalent to or even superior to the present state-of-the-art models in point cloud classification and object detection tasks. |
