【作者】 陈宇航； 师皓； 何静飞；

【Author】 CHEN Yuhang;SHI Hao;HE Jingfei;Radar Technology Research Institute,School of Information and Electronics,Beijing Institute of Technology;National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing;

【机构】 北京理工大学信息与电子学院雷达技术研究所； 天基智能信息处理全国重点实验室；

【摘要】 近年来随着机器学习、人工智能和图像生成技术的不断发展,3D点云生成系统已经成为人们日常生活中不能缺少的一项技术,而基于扩散模型的3D点云生成系统则是利用深度学习技术来产生三维点云。从原理上来看,该系统是从二维高斯分布的噪声中抽样以生成初步的噪声图像,然后通过神经网络逆向扩散处理这些噪声,逐渐趋近于目标图像。在这个过程中,整个扩散可以认为类似于马尔可夫链模型,也就是可以将不同步的变换之间都看做是相互独立的,其中每个点根据转移概率从噪声状态逆向扩散至目标状态。该研究采用了一种基于单个GPU的3D对象生成替代方法,使用文本到图像的扩散模型首先创建单个二维生成图像,然后应用另一个扩散模型从图像中生成点云。最终的实验结果表明,该3D点云生成系统具有较好的完成度,采样速度也非常快,具有轻量化、能够较快生成目标点云图的特性。更多还原

【Abstract】 In recent years,with the continuous development of machine learning,artificial intelligence,and image generation technology,3D point cloud generation systems have become an indispensable technology in people’s daily lives.The 3D point cloud generation system based on diffusion models utilizes deep learning technology to produce 3D point clouds.From a principled perspective,the 3D point cloud generation system based on diffusion models initially samples noise from a two-dimensional Gaussian distribution to generate a preliminary noisy image.Then,through neural network-based reverse diffusion,it processes these noises,gradually approaching the target image.In this process,the entire diffusion can be considered analogous to a Markov chain model,where different transformations can be regarded as mutually independent.Each point undergoes reverse diffusion from a noisy state to a target state based on transition probabilities.This research uses an alternative method for 3D object generation using a single GPU.It employs a text-to-image diffusion model to firstly create a single 2D generated image and then applies another diffusion model to generate a point cloud from the image.The final experimental results show that this 3D point cloud generation system has good completion rates,extremely fast sampling speeds,and is characterized by its lightweight nature and ability to quickly generate target point cloud images.更多还原