# Multi-scale Feature Extraction

SAM2 weights (HieraDet image encoder only) based on the timm library, derived from Facebook's Hiera small model.

MSI-Net is a visual saliency model that predicts human gaze points on natural images through a context encoder-decoder network trained on eye-tracking data.

Hiera is an efficient hierarchical Transformer architecture that optimizes spatial bias characteristics through MAE training, significantly improving parameter utilization efficiency

FocalNet is an image classification model based on the focal modulation network, pretrained by the Microsoft team on the ImageNet-22k dataset.

ResNet152 is an image classification model based on deep residual learning, which solves the gradient vanishing problem in deep network training through residual connections.

ResNet18 is an image classification model based on deep residual learning, which solves the difficulty of training deep networks through residual connections.

ResNet50 is an image classification model based on deep residual learning, which solves the vanishing gradient problem in deep neural networks through residual connections.