CAR-SAM: Cross-Attention Reconstruction for Post-Training Quantization of the Segment Anything Model

TL;DR

CAR-SAM introduces a novel post-training quantization framework for Segment Anything Models, effectively addressing attention dissipation and reconstruction oscillation to enable 4-bit model deployment.

Contribution

It proposes MAC and JCAR strategies specifically designed for SAMs, improving quantization stability and performance over existing methods.

Findings

Quantizes SAM models to 4-bit with significant accuracy gains

Outperforms existing PTQ methods by 14.6% and 6.6% mAP on SAM-B and SAM-L

Enhances model deployment on resource-constrained devices

Abstract

Segment Anything Models (SAMs) are extensively used in computer vision for universal image segmentation, but deploying them on resource-constrained devices is challenging due to their high computational and memory demands. Post-Training Quantization (PTQ) is a widely used technique for model compression and acceleration. However, existing PTQ methods fail to consider the cross-attention architecture in the SAM decoder. This degradation primarily stems from the unique challenges posed by SAMs: (1) Attention dissipation, where the attention information in the decoder, which is crucial for representing segmentation masks, collapses into a diffuse and non-semantic form under low-bit quantization; and (2) Reconstruction oscillation, where bidirectional coupling within the two-way transformer introduces cross-branch error interference and destabilizes convergence. To tackle these issues, we propose CAR-SAM, a unified quantization framework tailored for SAMs. Firstly, to mitigate attention dissipation, we introduce MatMul-Aware Compensation (MAC) mechanism that transfers activation-induced quantization errors from MatMul to preceding linear weights. Secondly, to mitigate oscillation in decoder optimization, we develop a Joint Cross-Attention Reconstruction (JCAR) strategy that jointly reconstructs coupled attention branches, suppressing oscillatory behavior and promoting stable convergence. Extensive experiments show that CAR-SAM robustly quantizes SAM models down to 4-bit precision, surpassing existing methods by 14.6% and 6.6% mAP on SAM-B and SAM-L respectively.