Abstract:Optical flow plays a crucial role in computer vision, and with the rapid advancements of Transformer-based models in natural language processing (NLP) and various visual tasks, Transformer architectures are becoming increasingly prevalent in the field of optical flow. In recent years, Transformer-based optical flow techniques have demonstrated superior accuracy compared to conventional CNN-based models. However, despite the existence of survey work on traditional and CNN-based methods for flow estimation, there is a lack of research focused exclusively on the emerging trend of Transformer-based approaches. This gap in the literature hinders the understanding and further development of these state-of-the-art techniques. To address this issue, we provide a systematic review and assessment of optical flow techniques with a sole focus on Transformer-based models, while also including CNN-based models for a fair comparison. We illustrate groundbreaking research on optical flow conducted in recent years using a timeline and establish a framework to comprehend the progression of Transformer implementation. Furthermore, we provide a detailed comparison of the performance of these techniques, experimental configurations, and computational efficiency, and emphasize visual comparisons to highlight the drawbacks and benefits of the benchmark datasets (FlyingChairs, Sintel, and KITTI2015). Our analysis reveals that while Transformer-based models achieve higher accuracy, they often come with increased computational demands compared to CNN-based approaches. This survey aims to bridge the gap in the existing literature and provide a comprehensive understanding of Transformer-based optical flow techniques, facilitating further research and development in this field.

Optical flow modeling and computation: A survey

Optical flow and scene flow estimation: A survey

Visual Optical Flow Computing: Algorithms and Applications

A Database and Evaluation Methodology for Optical Flow

Evolution of transformer-based optical flow estimation techniques: a survey

Survey on Unsupervised Learning Methods for Optical Flow Estimation

Optical Flow Dataset and Benchmark for Visual Crowd Analysis

A Quantitative Analysis of Current Practices in Optical Flow Estimation and the Principles Behind Them

Traditional and modern strategies for optical flow: an investigation

Research on Traditional and Deep Learning Strategies Based on Optical Flow Estimation - a Review

Optical Flow Estimation with Deep Learning, a Survey on Recent Advances

Review on Optical Flow Vector Fields Estimation Algorithm

Robust Optical Flow Computation: A Higher-Order Differential Approach

The Analysis and Development of Computed 2D Optical Flow Velocity Based on Phantom Data

Optical Flow for Video Super-Resolution: A Survey

On Benchmarking Optical Flow

Analysis on the Optimization of Optical Flow Estimation Algorithms

Optical Flow in Mostly Rigid Scenes

A Survey of Optical Flow Techniques for Robotics Navigation Applications

Determining Optical Flow under Space-Varying or Time-Varying Illumination

A Filtering-Based Framework for Optical Flow Estimation