Towards Partial Supervision for Generic Object Counting in Natural Scenes.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Generic object counting in natural scenes is complex, with current methods requiring extensive instance-level or absolute count supervision.
• High supervision levels limit the scalability and applicability of existing object counting techniques.

Purpose of the Study:

• To introduce a partially supervised setting that substantially lowers the supervision requirements for generic object counting.
• To propose and evaluate novel frameworks, lower-count (LC) and reduced lower-count (RLC), for efficient object counting.

Main Methods:

• Developed a dual-branch architecture integrating image classification and density estimation.
• Introduced the LC framework using only lower-count supervision for all categories.
• Proposed the RLC framework, extending LC with a weight modulation layer and category-independent density prediction, using lower-count supervision for a subset of categories and class-labels for others.

Main Results:

• The LC and RLC frameworks achieve performance comparable to state-of-the-art methods that use higher levels of supervision.
• Demonstrated the utility of the LC supervised density map for image-level supervised instance segmentation.
• Experiments conducted on COCO, Visual Genome, and PASCAL 2007 datasets validate the proposed approaches.

Conclusions:

• Partially supervised learning, specifically the LC and RLC frameworks, offers a viable and efficient alternative for generic object counting.
• The proposed methods significantly reduce annotation costs without compromising counting accuracy.
• The developed techniques show promise for broader applications, including instance segmentation.