Res2Net: A New Multi-Scale Backbone Architecture.

Area of Science:

• Computer Vision
• Deep Learning
• Image Recognition

Background:

• Multi-scale feature representation is crucial for computer vision tasks.
• Existing convolutional neural networks (CNNs) often represent features layer-wise.
• Advances in CNN backbones show improved multi-scale representation capabilities.

Purpose of the Study:

• To propose a novel building block for CNNs named Res2Net.
• To enhance multi-scale feature representation at a granular level within a single residual block.
• To improve the receptive field range of network layers.

Main Methods:

• Introduced hierarchical residual-like connections within a single residual block.
• Developed the Res2Net building block for CNNs.
• Integrated Res2Net into state-of-the-art CNN models like ResNet, ResNeXt, and DLA.

Main Results:

• Demonstrated consistent performance gains over baseline models on CIFAR-100 and ImageNet datasets.
• Verified the superiority of Res2Net on object detection, class activation mapping, and salient object detection.
• Showcased improved multi-scale feature representation at a granular level.

Conclusions:

• Res2Net offers a superior method for multi-scale feature representation in CNNs.
• The Res2Net block can be seamlessly integrated into existing CNN architectures.
• The proposed method achieves state-of-the-art results across various computer vision applications.