Final Year Project

Project Title

Comparative Analysis of Temporal Attention Methods for Image to Video Transfer of Frozen Vision Transformers

Description

This research project focuses on the theoretical computer vision domain and aims to study various temporal learning methods of Vision Transformers (ViTs) for video action recognition. The outcome will comprise a detailed comparative investigation of the temporal learning methods based on accuracy, time, and space complexity metrics. The methods will be analyzed using multiple training schemes with varying hyperparameters to assess their robustness. This study will be useful for the computer vision community to identify the best model under certain training settings and select a specific variant of a ViT for video recognition tasks based on their requirements and available resources.