Neural Network Density Maps for Object Counting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision systems face challenges in accurately performing object counting and instance segmentation, especially in natural scenes with occlusions, large intra-class variations, and diverse object categories, due to requirements for user-intensive annotations and difficulties in handling occlusions and background clutter.
Innovation Solution
A neural network architecture trained using image-level supervision techniques that construct density maps to predict object counts and locations, reducing the need for instance-level annotations and leveraging the subitizing range principle to minimize annotation burden, with a loss function optimizing both image classification and density prediction branches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If instance-level supervision techniques are used to train object counting and instance segmentation systems, then measurement precision is improved, but device complexity and annotation burden increase significantly
Solution Approach 1:
The patent uses image-level annotations as a simplified copy or approximation of instance-level annotations. Instead of requiring detailed instance-level labels, the system trains on coarser image-level count annotations, which are easier to obtain and reduce annotation complexity while still achieving accurate object counting and instance segmentation performance
Solution Approach 2:
The patent changes the annotation parameter from instance-level detailed labels to image-level count labels. This parameter change reduces the complexity and burden of data annotation while maintaining the system's ability to perform accurate object counting and instance segmentation through appropriate neural network architecture design
2Measurement precision
If traditional object detection methods are used to handle occluded objects, then ease of operation is maintained, but measurement precision deteriorates due to inability to accurately count occluded objects
Solution Approach 1:
The patent employs density maps as an intermediate representation that transforms the discrete object detection problem into a continuous density estimation problem. This dimensional change allows the system to count occluded objects by integrating density information across the image, overcoming the limitations of traditional detection methods that struggle with occluded instances
3Measurement precision
If neural networks are trained with full instance-level annotations, then object counting accuracy is improved, but loss of time increases due to extensive annotation requirements
Solution Approach 1:
The patent uses image-level annotations as a time-efficient proxy for instance-level annotations. By training on this simplified annotation format that requires significantly less time to create, the system achieves comparable instance segmentation accuracy without the extensive annotation time burden
Solution Approach 2:
The patent applies partial supervision by using only image-level count annotations rather than full instance-level annotations. This partial action approach reduces annotation time significantly while still providing sufficient training signal for the neural network to learn accurate object counting and instance segmentation
Data Source
AI summary
This disclosure relates to improved techniques for performing computer vision functions including common object counting and instance segmentation. The techniques described herein utilize a neural network architecture to perform these functions. The neural network architecture can be trained using image-level supervision techniques that utilize a loss function to jointly train an image classification branch and a density branch of the neural network architecture. The neural network architecture constructs per-category density maps that can be used to generate analysis information comprising global object counts and locations of objects in images.