Divide and Count: Generic Object Counting by Image Divisions.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object counting is crucial for various applications but often relies on category-specific models or detailed local annotations.
• Existing methods struggle with general object counting due to the need for prior knowledge or extensive labeling.

Purpose of the Study:

• To develop a general object counting method that does not require prior category information.
• To enable accurate global image-level counts solely from local image divisions without local annotations.

Main Methods:

• The proposed method divides images into local regions (region proposals or grid cells).
• An end-to-end deep learning architecture predicts global counts from these local divisions.
• A novel counting layer, based on the inclusion-exclusion principle, ensures count consistency across overlapping regions.

Main Results:

• The method was analyzed on Pascal-VOC2007.
• Evaluated on the large-scale MS-COCO generic object dataset.
• Validated on class-specific datasets: UCSD pedestrian, CARPK, and PUCPR+ car datasets.

Conclusions:

• The developed method offers a generalizable approach to object counting.
• It effectively predicts global counts from local image information without category priors.
• The technique demonstrates robust performance across diverse datasets.