Deep Learning over 3D Point Clouds

Speaker:  Ajmal Saeed Mian – Crawley, WN, Australia
Topic(s):  Artificial Intelligence, Machine Learning, Computer Vision, Natural language processing


3D point clouds are becoming an important data source for vision tasks such as auton-omous driving and robotic perception. However, deep learning over unstructured point clouds is challenging. We propose a spherical convolution kernel combined with octree guided neural net-work architecture for deep learning from unstructured point clouds. Spherical kernels systemati-cally quantize point neighborhoods to identify local geometric structures and avert dynamic ker-nel generation during network training for efficiency. We incorporate fuzzy mechanism into our discrete spherical convolutional kernel to avoid boundary effects during learning and varying point density during inference. We also propose an efficient graph convolutional network SegGCN that exploits ResNet like encoder blocks and 1x1 convolutions in the decoder. Next, we release Picasso, a CUDA-based library for deep learning over complex real-world 3D meshes. Picasso con-tains CUDA implementations of our Spherical Kernel, Fuzzy Spherical Kernel, three additional con-volution kernels (vertex2facet, facet2vertex, facet2facet) designed for 3D meshes and various network architecture design tools such as pooling, unpooling, residual layers, separable convolu-tion etc. Picasso also contains the first CUDA-based on-the-fly mesh simplification algorithm to facilitate hierarchical deep learning. We show the effectiveness of our novel convolution kernels, network architectures and mesh simplification algorithm on synthetic and real-world datasets. Finally, I will discuss our data capture hardware, challenges of outdoor LiDAR data and our open source annotation tool that can be used to generate large scale datasets for deep learning. Our methods are published in CVPR and PAMI. 

About this Lecture

Number of Slides:  55
Duration:  45 minutes
Languages Available:  English
Last Updated: 

Request this Lecture

To request this particular lecture, please complete this online form.

Request a Tour

To request a tour with this speaker, please complete this online form.

All requests will be sent to ACM headquarters for review.