BIASFit Point Cloud Segmentation for Noisy Geometric Primitive Fitting

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for fitting multi-model multi-instance geometric primitives in noisy point clouds, such as RANSAC-based approaches, often result in inferior fitting due to noisy inputs and cluttered scenes, which impede robustness and accuracy.

Innovation Solution

The BIASFit framework employs a fully convolutional neural network for point-wise segmentation of point clouds into primitive classes, with instance-aware boundary detection and geometric verification to correct misclassifications, using simulated range images and real-world data for training and testing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If RANSAC-based method is used for geometric primitive fitting, then the method can handle multi-model multi-instance scenes, but the fitting results are inferior due to noisy points and cluttered scenes

Engineering Contradiction:

Improvefitting accuracyVSAvoidnoisy points and cluttered scenes

Core Design Contradiction:

ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent applies segmentation by dividing the point cloud into multiple instances of geometric primitives through iterative RANSAC fitting. Each primitive instance is segmented as a separate hypothesis, allowing the system to handle multi-model multi-instance scenes by treating each instance independently rather than as a cluttered whole.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent uses preliminary action by generating primitive hypotheses before final fitting verification. The system first proposes potential primitive instances based on point cloud characteristics, then verifies these hypotheses through geometric constraints and error metrics, correcting misclassifications in a preliminary verification stage before final output.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If Hough Transform or global energy minimization is used, then geometric primitives can be detected, but the method suffers from noisy inputs and cluttered scenes similarly to RANSAC

Engineering Contradiction:

Improveprimitive detection accuracyVSAvoidnoisy inputs

Core Design Contradiction:

Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent implements feedback through an iterative verification process where primitive hypotheses are generated, then evaluated against geometric constraints and error metrics. The system provides feedback by identifying misclassified points and correcting them through re-fitting, continuously improving the detection accuracy by feeding back verification results into the fitting process.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent replaces traditional mechanical voting-based methods (Hough Transform) with a hypothesis-verification framework that uses geometric error metrics and probabilistic reasoning. Instead of accumulating votes in parameter space, the system substitutes this mechanical approach with statistical evaluation of primitive hypotheses against observed point cloud data.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Extent of automation

If deep neural networks are used for point cloud segmentation, then point-wisely segmentation into multiple classes can be achieved, but the system requires geometric verification to correct misclassification

Engineering Contradiction:

Improveautomatic segmentationVSAvoidclassification accuracy

Core Design Contradiction:

Extent of automationVSReliability

Solution Approach 1:

The patent uses geometric verification as an intermediary between deep neural network segmentation and final primitive classification. The neural network provides initial automated segmentation, then the geometric verification intermediary corrects misclassifications by evaluating geometric consistency, acting as a mediator that bridges the gap between automated segmentation and reliable classification.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent applies parameter changes by adjusting classification decisions based on geometric error metrics. When the neural network's initial classification shows high geometric error, the system changes the classification parameters by re-fitting primitives and re-evaluating points, dynamically adjusting classifications to improve reliability while maintaining automation.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20190279368A1Method and Apparatus for Multi-Model Primitive Fitting based on Deep Geometric Boundary and Instance Aware Segmentation

Publication Date: 2019.09.12 MITSUBISHI ELECTRIC RESEARCH LABORATORIES INC

AI summary

An image processing system includes an interface to transmit and receive data via a network, a processor connected to the interface, a memory storing an image processing program modules executable by the processor, wherein the image processing program causes the processor to perform operations. The operations include providing a point cloud of an image including objects into a segmentation network, segmenting point-wisely the point cloud into multiple classes of the objects and detecting boundaries of the objects using the segmentation network, wherein the segmentation network outputs a probability of associating primitive classes of the objects based on the segmented multiple classes and the segmented boundaries, verifying and refining the segmented multiple classes and the segmented boundaries using a predetermined fitting method, and correcting misclassification of the multiple classes of the objects by fitting the primitives to the multiple classes.