SIMD AdaBoost Decision Tree Traversal for Faster Object Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AdaBoost classifiers, particularly those using decision trees, are inefficient when implemented on conventional processors, leading to suboptimal performance in tasks like pedestrian detection due to the lack of effective utilization of parallel processing capabilities.
Innovation Solution
Implementing decision tree-based AdaBoost classification on single-instruction multiple-data (SIMD) processors, where features are packed into vectors for simultaneous processing, optimizing the traversal of decision trees and feature extraction across multiple scales using a sliding window approach.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If decision tree-based AdaBoost classification is implemented on conventional processors, then the classifier can be executed, but the processing efficiency is low and computational time is excessive
Solution Approach 1:
The patent segments the decision tree traversal process into independent parallel tasks that can be executed simultaneously on SIMD processor elements. Each processor element handles a separate data instance or feature subset, dividing the monolithic sequential traversal into concurrent parallel operations that eliminate the bottleneck of conventional sequential processing
Solution Approach 2:
The patent transitions from sequential single-data processing to parallel multi-data processing by utilizing the SIMD architecture's inherent vector processing capabilities. Multiple data instances are processed simultaneously across multiple processor elements, adding a parallelism dimension that transforms the computational approach from serial to concurrent execution
2Adaptability or versatility
If features are processed sequentially in traditional AdaBoost implementation, then memory access is simplified, but the utilization of parallel processing capabilities is insufficient
Solution Approach 1:
The patent merges multiple feature values into packed vector structures that can be processed simultaneously by SIMD instructions. Features from multiple data instances are combined into single memory structures, allowing parallel access and processing across multiple processor elements while maintaining memory efficiency through compact packed representations
Solution Approach 2:
The patent changes the data representation parameters by organizing features into vector-packed formats with specific alignment and grouping. This parameter transformation enables the data structure to be compatible with SIMD instruction sets, allowing parallel processing operations to be applied efficiently across multiple features simultaneously
3Speed
If multiple object patches are evaluated sequentially, then the implementation is simpler on conventional processors, but the detection speed is reduced
Solution Approach 1:
The patent segments the object patch evaluation process into independent parallel tasks assigned to different SIMD processor elements. Each processor element independently evaluates a subset of object patches simultaneously, dividing the sequential evaluation bottleneck into concurrent parallel operations that multiply the detection throughput
Solution Approach 2:
The patent performs preliminary packing and organization of object patch features into vector structures before the actual classification evaluation. This preliminary preparation arranges data in a format optimized for parallel SIMD processing, enabling subsequent concurrent evaluation to proceed efficiently without runtime data reorganization overhead
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
In described examples of a method for object classification in a decision tree based adaptive boosting (AdaBoost) classifier implemented on a single-instruction multiple-data (SIMD) processor, the method includes receiving (700) feature vectors extracted from N consecutive window positions in an image in a memory coupled to the SIMD processor and evaluating (708) the N consecutive window positions concurrently by the AdaBoost classifier using the feature vectors and vector instructions of the SIMD processor, in which the AdaBoost classifier concurrently traverses (714) decision trees for the N consecutive window positions until classification is complete (712) for the N consecutive window positions.