Object Detection via Selective Pooling for Overlapping Features

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object detection methods face challenges in efficiently extracting features based on object locations, leading to suboptimal performance in identifying and classifying objects, especially when objects overlap.

Innovation Solution

The proposed solution involves a two-stage object detection framework that uses a Convolutional Neural Network (CNN) to convert images into feature maps, followed by a Region Proposal Network (RPN) to generate regions of interest. A selective pooling layer is applied to the feature bank to extract target feature maps, which are then used to determine object classes and boundaries through a fully connected layer.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional object detection methods are used, then object detection can be performed, but detection performance and accuracy are suboptimal, especially for overlapping objects

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent divides the feature map into multiple channels, each channel processing specific spatial regions or feature types independently. This segmentation allows parallel processing of different object regions, improving both detection accuracy for overlapping objects and computational efficiency through channel-wise parallelization

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different processing strategies to different spatial regions and channels based on their specific characteristics. Each channel can be optimized for specific types of objects or regions, allowing local optimization that improves overall detection performance while maintaining computational efficiency

Inventive Principle:
Principle #3Local quality

2Productivity

If feature map dimensionality is reduced, then computational efficiency improves, but feature information may be lost

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidfeature information loss
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent transforms the feature representation by organizing features into multiple channels, effectively changing the dimensional organization from a single large feature map to a structured multi-channel format. This dimensional reorganization reduces computational complexity while preserving feature information through the channel structure that maintains spatial and semantic relationships

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

Data Source

PatentEP3649574B1Object detection based on deep neural network
Publication Date: 2025.02.19 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3649574B1 patent drawingFigure 1
  • EP3649574B1 patent drawingFigure 2
  • EP3649574B1 patent drawingFigure 3

AI summary

Implementations of the subject matter described herein relate to object detection based on deep neural network. With a given input image, it is desired to determine a class and a boundary of one or more objects within the input image. Specifically, a plurality of channel groups is generated from a feature map of an image, the image including at least a region corresponding to a first grid. A target feature map is extracted from at least one of the plurality of channel groups associated with a cell of the first grid. Information related to an object within the region is determined based on the target feature map. The information related to the object may be a class and/or a boundary of the object.