Probabilistic Boosting Tree GPU Evaluation via Stack Implementation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Discriminative classifiers, such as Probabilistic Boosting Trees, are inefficient in 3D pose detection due to recursive calls, which hinder performance in real-time and medical applications, necessitating improved efficiency methods.

Innovation Solution

Implementing a stack-based evaluation method for Probabilistic Boosting Trees on a GPU, replacing recursive calls with a stack implementation to accelerate training and evaluation, utilizing CUDA for parallel processing and efficient memory access.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If recursive calls are used in Probabilistic Boosting Trees for classification, then the classifier can make fuzzy decisions at internal nodes, but the detection speed decreases and CPU efficiency is reduced

Engineering Contradiction:
Improveclassification accuracyVSAvoiddetection speed
Core Design Contradiction:
ReliabilityVSSpeed

Solution Approach 1:

The patent replaces the traditional recursive call mechanism (mechanical system) with an iterative approach using explicit stack data structures. This substitution eliminates the overhead of recursive function calls while maintaining the same classification logic, thereby improving detection speed without sacrificing classification accuracy

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent transitions from a single-threaded recursive evaluation to a multi-threaded parallel evaluation model. By organizing the tree traversal in a way that enables parallel processing across multiple threads, the system achieves both high accuracy through complete tree evaluation and high speed through parallel computation

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Reliability

If Probabilistic Boosting Trees are evaluated at every pixel in an image, then comprehensive object detection is achieved, but the computational bottleneck increases processing time

Engineering Contradiction:
Improvedetection completenessVSAvoidprocessing throughput
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent divides the image processing task into independent pixel-level segments that can be processed in parallel. Each pixel evaluation is treated as an independent unit of work that can be assigned to different threads, enabling comprehensive detection across all pixels while maintaining high processing throughput through parallelization

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary preparation of the Probabilistic Boosting Tree structure and feature data before actual pixel evaluation. By pre-computing and organizing the classification tree and feature representations, the system reduces the computational burden during the actual pixel-wise evaluation, thereby improving processing throughput while maintaining detection completeness

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP2434435B1Method and system for evaluation using probabilistic boosting trees
Publication Date: 2018.10.31 SIEMENS CORP
  • EP2434435B1 patent drawingFigure 1
  • EP2434435B1 patent drawingFigure 2
  • EP2434435B1 patent drawingFigure 3

AI summary

A method and system for evaluating probabilistic boosting trees is disclosed. In an embodiment, input data is received at a graphics processing unit. A weighted empirical distribution associated with each node of the probabilistic boosting tree is determined using a stack implementation. The weighted empirical distribution associated with each node is added to a total posterior distribution value.