Involution - Inverting the Inherence of Convolution for Visual Recognition

Explore a comprehensive analysis of the research paper "Involution: Inverting the Inherence of Convolution for Visual Recognition" in this 31-minute video. Delve into the innovative Involution Operator, which challenges traditional convolutional neural network principles by introducing spatial-specific and channel-agnostic computations. Learn how this novel approach compares to classic convolutions and local self-attention architectures, and discover its potential to improve performance while reducing computational costs in various computer vision tasks. Gain insights into the experimental results across ImageNet classification, COCO detection and segmentation, and Cityscapes segmentation benchmarks.

- Intro & Overview
- Principles of Convolution
- Towards spatial-specific computations
- The Involution Operator
- Comparison to Self-Attention
- Experimental Results
- Comments & Conclusion