Grid-Based Point Cloud Segmentation Without Bounding Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing point cloud segmentation methods suffer from low efficiency and accuracy due to the need to remove redundant bounding boxes and rely on inaccurate bounding box predictions.
Innovation Solution
A point cloud segmentation method using a pre-trained neural network for direct point-by-point classification in a gridding scenario space, where points are assigned to target grids based on their instance category, reducing the need for multiple stages and minimizing computational loss.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If top-down segmentation method with multiple bounding boxes is used, then instance segmentation can be performed, but segmentation efficiency deteriorates due to redundant bounding box removal
Solution Approach 1:
The patent divides the scenario space into multiple grids and assigns each grid to a specific instance category. This grid-based segmentation eliminates the need for redundant bounding boxes by directly mapping points to instance categories through grid identification, thereby improving segmentation efficiency while maintaining instance segmentation capability.
Solution Approach 2:
The patent introduces a grid dimension to the scenario space, transforming the traditional bounding box-based approach into a grid-based classification system. By adding this spatial dimension, the system can directly classify points into instance categories without requiring multiple bounding box predictions and subsequent filtering.
2Reliability
If top-down segmentation method with bounding box prediction is used, then instance segmentation can be performed, but segmentation accuracy deteriorates due to dependency on bounding box prediction accuracy
Solution Approach 1:
By segmenting the scenario space into grids and assigning each grid to a specific instance category, the patent eliminates the error propagation from bounding box prediction. The neural network directly predicts grid categories for each point, bypassing the intermediate bounding box prediction stage that causes accuracy deterioration.
Solution Approach 2:
The patent extracts and removes the problematic bounding box prediction stage from the segmentation pipeline. By directly mapping points to grids and then to instance categories, the system eliminates the source of accuracy errors while preserving the essential instance segmentation functionality.
3Reliability
If multiple stages of bounding box prediction and filtering are used, then instance segmentation can be achieved, but computational overhead increases
Solution Approach 1:
The patent merges the instance segmentation functionality into a single grid-based classification stage. Instead of separate stages for bounding box prediction, filtering, and segmentation, the system combines these functions into one process where points are directly classified into instance categories through grid identification, reducing computational overhead.
Solution Approach 2:
The patent performs preliminary organization of the scenario space into grids with pre-assigned instance categories. This preliminary action eliminates the need for subsequent filtering and refinement stages, as the grid structure is already optimized for direct instance classification, thereby reducing overall computational complexity.
Data Source
AI summary
Provided are a point cloud segmentation method and apparatus, a device, and a storage medium. The point cloud segmentation method includes: acquiring a to-be-processed point cloud; obtaining, in a gridding scenario space to which respective point in the point cloud belongs, a target grid corresponding to the respective point through a pre-trained neural network, wherein the pre-trained neural network is obtained by training a sample point cloud and a sample target grid corresponding to the sample point cloud in a sample gridding scenario space; and outputting a point cloud corresponding to a respective instance according to an instance category corresponding to the target grid, wherein the same target grid has the same instance category.