Samhq model addition

[sushmanthreddy]

close #31137

Pull Request Title: Add HQ-SAM Functionality to Transformers Library

Model Overview

HQ-SAM (Segment Anything in High Quality) is an enhanced version of the Segment Anything Model (SAM), addressing limitations in mask quality for intricate structures and challenging segmentation tasks. The model refines SAM’s predictions using a High-Quality Output Token and Global-Local Feature Fusion while preserving SAM’s efficiency and zero-shot generalization capabilities.

According to the original implementation, HQ-SAM significantly improves mask boundaries and reduces segmentation errors by introducing minimal additional parameters (<0.5%) and computational overhead. The model is designed to maintain compatibility with SAM’s existing prompt-based design and mask decoder architecture.

Repository and Weights

The HQ-SAM implementation and pre-trained weights are available in the following repository:
https://github.com/SysCV/sam-hq

HQ-SAM provides three pre-trained weight variants:

• sam_hq_vit_b – Small vision encoder.
• sam_hq_vit_l – Medium vision encoder.
• sam_hq_vit_h – Large vision encoder.

The main difference between these variants is the size of the Vision Transformer (ViT) encoder, while the prompt encoder and mask decoder remain unchanged.

Functionality

For each input (e.g., bounding boxes, 2D points, or coarse masks), HQ-SAM predicts high-quality binary masks that enhance segmentation precision. Improvements include:

• More accurate boundaries.
• Correction of coarse masks and segmentation errors.
• Enhanced detail preservation for thin structures and complex object geometries.

Reviewers: @molbap

[sushmanthreddy]

still work in progress ,work is being converted from model files to modular file due to there is lot of code of code we can reuse from the sam model