Analysis of Mainstream Models Based on Deep Learning

Optical flow estimation is a hot topic that significantly impacts the development of the computer vision field, especially in pattern recognition. With advancements in artificial intelligence technologies, deep learning has seemingly opened up a new path for optical flow estimation. This paper focuses on four of the most influential deep learning-based models: the FlowNet series, the PWC-Net series, the RAFT series, and the SPyNet series. Each category includes representative models that showcase subsequent improvements. Firstly, this paper analyzes the neural convolutional network architecture and design of the basic models in each category. Following this, the optimization principles behind the improved versions of these models are explored. Finally, the strengths and weaknesses of these models by evaluating their Average Endpoint Error (AEE) results across various datasets are summarized. The paper concludes with an overall summary of the findings, providing insights into the progress and future directions of optical flow estimation in the context of deep learning.