Incremental Sparse 3D Convolution for Real-Time Point Cloud Processing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing incremental convolution computation methods are not suitable for sparse convolutions, which are widely used in 3D segmentation and detection tasks, particularly in applications like autonomous driving and robotics, due to their reliance on dense convolutions, leading to inefficiencies in processing 3D point cloud data.

Innovation Solution

A method utilizing incremental sparse 3D convolution, including an incremental submanifold sparse convolution layer and non-linear layer, to constrain active residual sites and propagate residuals efficiently, reducing computational complexity and enabling real-time 3D point cloud data processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If dense convolution methods are used for incremental computation, then computation speed can be improved through incremental updates, but the method is not suitable for sparse convolution tasks required in 3D segmentation and detection

Engineering Contradiction:
Improvecomputation speedVSAvoidapplicability to sparse convolution tasks
Core Design Contradiction:
ProductivityVSAdaptability or versatility

Solution Approach 1:

The patent segments the convolution operation into sparse and dense components, applying incremental computation only to the sparse portion while maintaining full computation for dense operations. This allows the method to handle sparse 3D point cloud data efficiently without compromising the applicability to sparse convolution tasks.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter representation by introducing a sparsity mask that identifies active regions in 3D space. By dynamically adjusting which regions undergo incremental updates versus full recomputation based on sparsity patterns, the method achieves both speed improvement and compatibility with sparse convolution requirements.

Inventive Principle:
Principle #35Parameter changes

2Ease of manufacture

If single-frame computation method is used for 3D point cloud data, then implementation is simple, but redundant computations occur due to duplicate information between frames

Engineering Contradiction:
Improveimplementation simplicityVSAvoidredundant computation
Core Design Contradiction:
Ease of manufactureVSLoss of energy

Solution Approach 1:

The patent performs preliminary identification of changed regions between frames by comparing point cloud data and generating a difference mask. This preliminary action allows subsequent incremental convolution to operate only on affected regions, eliminating redundant computations while maintaining implementation feasibility through modular design.

Inventive Principle:
Principle #10Preliminary action

3Speed

If incremental computation is applied to 3D point cloud data processing, then processing speed is improved, but computational complexity increases due to sparse convolution requirements

Engineering Contradiction:
Improveprocessing speedVSAvoidcomputational complexity
Core Design Contradiction:
SpeedVSDevice complexity

Solution Approach 1:

The patent extracts and isolates the sparse convolution operations from the overall processing pipeline, applying incremental computation specifically to these extracted sparse regions. By separating the computation into distinct sparse and dense handling paths, the method improves processing speed for sparse data without unnecessarily complicating the overall system.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4254362B1Three-dimensional point cloud data processing method and device based on incremental sparse three-dimensional convolution
Publication Date: 2026.03.18 TSINGHUA UNIVERSITY
  • EP4254362B1 patent drawingFigure 1
  • EP4254362B1 patent drawingFigure 2
  • EP4254362B1 patent drawingFigure 3~5

AI summary

In a method for processing three-dimensional (3D) point cloud data based on incremental sparse 3D convolution, a computer device obtains 3D point cloud data and forms a training set by processing the 3D point cloud data. The computer device constructs and trains a sparse 3D convolutional network model by inputting the training set. The computer device constructs an incremental sparse 3D convolutional network model by performing incremental replacement of sparse convolutional layers of the trained sparse 3D convolutional network model. The computer device inputs real-time 3D point cloud data into the incremental sparse 3D convolutional network model, and determines an output result as a result of processing the real-time 3D point cloud data. Processing of the 3D point cloud data at least includes 3D semantic segmentation, target detection, 3D classification and video processing.