Multi-perspective Object Detection via Feature Map Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computerized object detection systems are unable to effectively cross-reference images from multiple perspectives of a scene, limiting their ability to accurately detect and identify objects, unlike human analysts who can analyze images from different views to better identify positions and classify objects.
Innovation Solution
The introduction of a fusion layer that allows object detection pipelines to share feature map data across different perspectives, enabling the combination and cross-referencing of data from multiple perspectives to improve detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If object detection is performed on single-perspective images only, then the detection system remains simple and fast, but the accuracy of object detection and identification is limited
Solution Approach 1:
The patent merges multiple single-perspective detection pipelines into a unified multi-perspective detection system. Feature maps from different perspective images are combined through fusion layers, allowing the system to leverage information from multiple views simultaneously. This merging approach improves detection accuracy by providing more comprehensive object information while maintaining a structured detection framework.
Solution Approach 2:
The patent introduces a new dimension to object detection by incorporating multiple perspectives (views) of the same scene. Instead of analyzing objects from a single viewpoint, the system processes images from different angles and positions, adding a spatial perspective dimension. This enables the detection system to overcome limitations of single-view detection, such as occlusions and ambiguous orientations, thereby improving overall detection accuracy.
2Measurement precision
If multiple perspective images are processed independently, then processing speed is maintained, but the ability to cross-reference and accurately identify objects is lost
Solution Approach 1:
The patent combines feature maps from multiple independent detection pipelines through fusion layers. Instead of processing each perspective image completely separately and then comparing results, the system merges the extracted features at the feature map level, enabling direct cross-referencing of objects across different perspectives. This approach improves identification accuracy while avoiding the computational overhead of independent full-processing pipelines.
Solution Approach 2:
The patent introduces fusion layers as intermediary components between independent detection pipelines. These fusion layers act as mediators that receive feature maps from multiple perspectives, perform cross-referencing operations, and generate unified detection results. This intermediary approach enables efficient cross-perspective object identification without requiring complete reprocessing of all images, thus maintaining processing efficiency while improving accuracy.
3Measurement precision
If feature map data is shared across detection pipelines, then detection accuracy improves, but computational overhead and system complexity increase
Solution Approach 1:
The patent merges feature map data from multiple detection pipelines through fusion layers, enabling shared utilization of extracted features across different perspectives. By combining features at the map level rather than processing images independently, the system achieves improved multi-perspective detection accuracy while avoiding redundant computation. The fusion operations efficiently integrate information from multiple sources without requiring complete reprocessing of all input data.
Data Source
AI summary
Technology disclosed herein may involve a computing system that (i) generates (a) a first feature map based on a first visual input from a first perspective of a scene utilizing at least one first neural network and (b) a second feature map based on a second visual input from a second, different perspective of the scene utilizing at least one second neural network, where the first perspective and the second perspective share a common dimension, (ii) based on the first feature map and a portion of the second feature map corresponding to the common dimension, generates cross-referenced data for the first visual input, (iii) based on the second feature map and a portion of the first feature map corresponding to the common dimension, generates cross-referenced data for the second visual input, and (iv) based on the cross-referenced data, performs object detection on the scene.