Incremental sparse 3D convolutional computing architecture, system and device

Abstract

The invention provides an incremental sparse three-dimensional convolution three-dimensional point cloud data processing method and device. The method obtains three-dimensional point cloud data and performs data preprocessing to form a training set of three-dimensional point cloud data; constructs a sparse three-dimensional volume the sparse 3D convolutional network model, and input the training set of 3D point cloud data into the sparse 3D convolutional network model for model training; build an incremental sparse 3D volume based on the sparse 3D convolutional network model after training The cumulative network model; the real-time three-dimensional point cloud data is input into the incremental sparse three-dimensional convolution network model, and the output result is the processing result of the real-time three-dimensional point cloud data. Through the present invention, the online task using sparse convolution can be accelerated, the incremental calculation can be realized, and the network prediction can be accelerated.

Description

technical field [0001] The present invention relates to the technical field of deep learning, and in particular, to an incremental sparse three-dimensional convolution three-dimensional point cloud data processing method, a computer device and a non-transitory computer-readable storage medium. Background technique [0002] Incremental sparse 3D convolution computations have important applications in timing problems in computer vision. Some timing problems, such as visual tasks on video, usually use a single-frame calculation method to input each frame into a convolutional neural network to obtain a separate result, and the use of incremental computing technology can reduce repeated calculations, because the video frame There is a lot of duplicate information between, resulting in faster processing speed. At present, most of the existing incremental convolution calculation methods are for dense convolution operations, and it is difficult to apply to the recently emerging sub...

AI Technical Summary

Problems solved by technology

However, the existing incremental convolution calculations are all aimed at dense convolution operations and cannot be applied to sparse convolutions.

Images