针对点云补全中 CD 损失参数调优耗时的问题,提出梯度匹配损失蒸馏搜索无需调参的加权 CD,并通过双层优化训练,Landau CD 在多个基准上达到最先进水平。
摘要:
3D点云增强了机器人感知环境几何信息的能力,使抓取姿态检测和场景理解等许多下游任务成为可能。然而,这些任务的性能严重依赖输入数据的质量,因为不完整的数据会导致较差的结果和失败案例。近期为基于深度学习的点云补全设计的训练损失函数,例如 Chamfer 距离(CD)及其变体(如 HyperCD),表明良好的梯度加权方案可以显著提升性能。但是,这些基于 CD 的损失函数通常需要进行与数据相关的参数调优,对于数据量庞大的任务而言非常耗时。为解决这一问题,本文旨在寻找一族无需参数调优的加权训练损失(加权 CD)。为此,我们提出一种搜索方案——基于梯度匹配的损失蒸馏,通过模仿 HyperCD 与加权 CD 之间反向传播中的学习行为来寻找优良的候选损失函数。在此基础上,我们提出一种新的双层优化公式,以基于加权 CD 损失训练骨干网络。我们观察到:(1)借助适当的加权函数,加权 CD 始终能够取得与 HyperCD 相近的性能;(2)Landau 加权 CD(即 Landau CD)在点云补全上可以超越 HyperCD,并在多个基准数据集上取得新的最先进结果。我们的演示代码可在 https://github.com/Zhang-VISLab/IROS2024-LossDistillationWeightedCD 获取。
英文摘要:
3D point clouds enhanced the robot's ability to perceive the geometrical information of the environments, making it possible for many downstream tasks such as grasp pose detection and scene understanding. The performance of these tasks, though, heavily relies on the quality of data input, as incomplete can lead to poor results and failure cases. Recent training loss functions designed for deep learning-based point cloud completion, such as Chamfer distance (CD) and its variants (\eg HyperCD ), imply a good gradient weighting scheme can significantly boost performance. However, these CD-based loss functions usually require data-related parameter tuning, which can be time-consuming for data-extensive tasks. To address this issue, we aim to find a family of weighted training losses ({\em weighted CD}) that requires no parameter tuning. To this end, we propose a search scheme, {\em Loss Distillation via Gradient Matching}, to find good candidate loss functions by mimicking the learning behavior in backpropagation between HyperCD and weighted CD. Once this is done, we propose a novel bilevel optimization formula to train the backbone network based on the weighted CD loss. We observe that: (1) with proper weighted functions, the weighted CD can always achieve similar performance to HyperCD, and (2) the Landau weighted CD, namely {\em Landau CD}, can outperform HyperCD for point cloud completion and lead to new state-of-the-art results on several benchmark datasets. {\it Our demo code is available at \url{https://github.com/Zhang-VISLab/IROS2024-LossDistillationWeightedCD}.}
Density-aware Chamfer Distance as a Comprehensive Metric for Point Cloud Completion
SenseTime-CUHK Joint Lab, The Chinese University of Hong Kong(商汤科技-香港中文大学联合实验室,香港中文大学);S-Lab, Nanyang Technological University(南洋理工大学S-Lab);SenseTime Research(商汤科技研究院);Shanghai AI Laboratory(上海人工智能实验室);Centre of Perceptual and Interactive Intelligence(感知与交互智能中心)
GSCompleter: A Distillation-Free Plugin for Metric-Aware 3D Gaussian Splatting Completion in Seconds
School of Computer Science and Technology, East China Normal University(东华大学计算机科学与技术学院);Shanghai Innovation Institute(上海创新研究院);Chongqing Key Laboratory of Precision Optics, Chongqing Institute of East China Normal University(重庆精密光学重点实验室,东华大学重庆研究院);Shanghai Key Laboratory of Computer Software Evaluating and Testing(上海计算机软件评测测试重点实验室);Department of Computer Science and Engineering, Shanghai Jiao Tong University(上海交通大学计算机科学与工程学院)
Memory-Guided Point Cloud Completion for Dental Reconstruction
College of Information and Intelligent Science, Donghua University, Shanghai, China(信息与智能科学学院,东华大学,上海,中国);Institute of Artificial Intelligence, Shanghai University, Shanghai, China(人工智能研究院,上海大学,上海,中国)
Point-Based 3D Reconstruction from Sparse Views under Known Illumination
Aalborg University(奥尔堡大学);Centre for Software Technology, University of Southern Denmark(南丹麦大学软件技术中心);Pioneer Centre for Artificial Intelligence(先锋人工智能中心);Technical University of Denmark(丹麦技术大学);Technische Universität Berlin(柏林工业大学)
A Conditional Point Diffusion-Refinement Paradigm for 3D Point Cloud Completion
The Chinese University of Hong Kong(香港中文大学);Shanghai AI Laboratory(上海人工智能实验室);Nanyang Technological University(南洋理工大学);University of California, San Diego(加利福尼亚大学圣迭戈分校)
Three-dimentional reconstruction of complex, dynamic population canopy architecture for crops with a novel point cloud completion model: A case study in Brassica napus rapeseed
SurgPointTransformer: Vertebrae Shape Completion with RGB-D Data
Research in Orthopedic Computer Science, University Hospital Balgrist, University of Zürich(苏黎世大学、巴尔格里斯特大学医院骨科计算机科学研究所);Balgrist University Hospital, University of Zurich(苏黎世大学巴尔格里斯特大学医院)
APML: Adaptive Probabilistic Matching Loss for Robust 3D Point Cloud Reconstruction
Center for Machine Vision and Signal Analysis (CMVS), University of Oulu (Finland)(机器视觉与信号分析中心(CMVS)、奥卢大学(芬兰));Okinawa Institute of Science and Technology (OIST), Japan(冲绳科学技术研究所(OIST)、日本)
3D Brain Reconstruction by Hierarchical Shape-Perception Network from a Single Incomplete Image
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences(中国科学院深圳先进技术研究院);University of Chinese Academy of Sciences(中国科学院大学);School of Biomedical Engineering, Shenzhen University(深圳大学生物医学工程学院);Gaoling School of Artificial Intelligence, Renmin University of China(中国人民大学高瓴人工智能学院);College of Computer Science and Technology, Qingdao University(青岛大学计算机科学与技术学院)
Sparse Depth-Guided Attention for Accurate Depth Completion: A Stereo-Assisted Monitored Distillation Approach
Huawei Cloud Computing Technology Co., Ltd.(华为云计算技术有限公司);Shenzhen Institute of Artificial Intelligence and Robotics for Society(深圳人工智能与机器人学会);Institute of Robotics and Intelligent Manufacturing, The Chinese University of Hong Kong, Shenzhen(香港中文大学(深圳)机器人与智能制造研究院)
Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy
Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences(中国科学院香港创新研究院);Institute of Automation, Chinese Academy of Sciences(中国科学院自动化研究所);University of Chinese Academy of Sciences(中国科学院大学)
L2COcc: Lightweight Camera-Centric Semantic Scene Completion via Distillation of LiDAR Model
Institute of Cyber-Systems and Control, Zhejiang University(浙江大学网络系统与控制研究所);School of Computation, Information and Technology, Technical University of Munich(慕尼黑工业大学计算、信息与技术学院);Zhejiang Guoli Security Technology Co., Ltd.(浙江国利安全科技有限公司)
DistillGrasp: Integrating Features Correlation with Knowledge Distillation for Depth Completion of Transparent Objects
College of Computer Science and Technology, Guangdong University of Technology(广东工业大学计算机科学与技术学院);Hasselt University(哈塞尔特大学);Flanders Make(佛兰德制造机构);Zhongguancun Laboratory(中关村实验室)
PointSFDA: Source-free Domain Adaptation for Point Cloud Completion
School of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics(南京航空航天大学计算机科学与技术学院);Delft University of Technology(代尔夫特理工大学);School of Nursing, The Hong Kong Polytechnic University(香港理工大学护理学院)