Asymmetric Convolution: An Efficient and Generalized Method to Fuse Feature Maps in Multiple Vision Tasks.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Feature fusion is crucial for computer vision tasks.
• Current methods include parameter-free and learnable approaches.
• Both existing methods have limitations in performance and efficiency, especially with diverse feature shapes.

Purpose of the Study:

• To address the limitations of existing feature fusion methods.
• To propose a novel, generalized module for efficient and effective feature fusion.
• To enhance the performance of computer vision models through improved feature integration.

Main Methods:

• An in-depth analysis of parameter-free and learnable fusion techniques was conducted.
• A generalized Asymmetric Convolution Module (ACM) was developed.
• A mathematically equivalent, efficient method for fusing features of different shapes was proposed.

Main Results:

• The Asymmetric Convolution Module (ACM) demonstrated efficient fusion of feature maps with varying shapes.
• ACM effectively fuses multiple features of different types, unlike parameter-free methods.
• Integration of ACM into state-of-the-art models yielded significant performance improvements across three vision tasks.

Conclusions:

• The proposed Asymmetric Convolution Module (ACM) offers a flexible and efficient solution for feature fusion in computer vision.
• ACM overcomes limitations of existing methods, enabling better performance and reduced computational load.
• ACM shows broad applicability and significant potential for advancing various computer vision applications.