Layered insights into Pyramid feature fusion architecture for SSL
Keywords: Machine Learning, Artificial Intelligence, Computer Vision, Self-supervised learning
TL;DR: A novel SSL approach for Gaussian pyramid architectures and the contribution of each layer into the predictive power
Abstract: In this paper, we introduce a novel self-supervised learning(SSL) approach leveraging pyramid layers to extract essential visual features. Employing image inpainting as the pretext task, we empirically demonstrate the effectiveness of this methodology by rigorously evaluating the trained network's performance in downstream image classification/segmentation tasks. Our findings underscore the substantial performance improvement achieved through SSL. Furthermore, we study the individual contributions of each self-supervised trained feature pyramid layer to the model's performance enhancement. Findings here, led us to conclude that optimal number and size of feature pyramid layers vary for each model and significantly influence overall performance.
Submission Number: 205
Loading