Informative Point cloud Dataset Extraction for Classification via Gradient-based Points Moving

Wenxiao Zhang, Ziqi Wang, Li Xu, Xun Yang, Jun Liu

Abstract: Point cloud plays a significant role in recent learning-based vision tasks, which contain additional information about the physical space compared to 2D images. However, such a 3D data format also results in more expensive computation costs to train a sophisticated network with large 3D datasets. Previous methods for point cloud compression focus on compacting the representation of each point cloud for better storage and transmission. In this paper, we introduce a new open problem in point cloud field: Can we compress a large point cloud dataset into a much smaller synthetic dataset while preserving the important information of the original large dataset?} In other words, we explore the possibility of training a network on a smaller dataset of informative point clouds extracted from the original large dataset but maintaining similar network performance. Training on this small synthetic dataset could largely improve the training efficiency. To explore this new open problem, we formulate it as a parameter-matching issue where a network could get similar network parameters after training on the original set and the generated synthetic set, respectively. We find that we could achieve this goal by moving the critical points within each initial point cloud through an iterative gradient matching strategy. We conduct extensive experiments on various synthetic and real-scanned 3D object classification benchmarks, showing that our synthetic dataset could achieve almost the same performance with only 5\% point clouds of ScanObjectNN dataset compared to training with the full dataset.

Primary Subject Area: [Experience] Multimedia Applications

Secondary Subject Area: [Generation] Generative Multimedia

Relevance To Conference:

In this paper, we introduce a new open task in the point cloud field, in which we explore the possibility of training a network on a smaller dataset of informative point clouds extracted from the original large dataset but maintaining similar network performance. Our work is tightly related to 3D modal data understanding and processing which plays an important role in multimedia area. There are several previous point cloud classification related papers published in ACM MM conference:

[1] Object Point Cloud Classification via Poly-Convolutional Architecture Search. MM'21
[2] SUG: Single-dataset Unified Generalization for 3D Point Cloud Classification. MM'23

Submission Number: 1822