ACDC: Adaptive Cloud-Device Collaboration for Efficient and Accurate Semantic Segmentation

Keming Ye, Zhou Zhao, Fan Wu, Shengyu Zhang

Keywords: Cloud-Device Collaboration, Semantic Segmentation, Uncertainty Quantification

Abstract: Semantic segmentation is a vital task in computer vision with wide-ranging practical applications. Advanced segmentation models achieve high accuracy but face computational challenges for device-based deployment, while lightweight models often produce coarse predictions, potentially losing pixel-level details. To address these limitations, this paper introduces the Adaptive Cloud-Device Collaboration (ACDC) framework, which combines the efficiency of device-side models with the robust capabilities of cloud-side models. ACDC employs an adaptive uncertainty detection mechanism to capture pixel-wise distributional shifts, filtering challenging samples for fine-grained processing on the cloud, and fuses dual-granularity predictions to achieve precise results. The framework comprises three key modules: Device-Aware Adaptive Segmentor (DAS) for coarse segmentation and uncertainty detection using a two-stage Uncertainty Decoupler; Dynamic Cloud Augmentation Module (DCAM) for processing challenging samples and adaptive update; and Collaborative Fusion Engine (CFE) for dual-granularity integration. Extensive experiments demonstrate that ACDC improves segmentation accuracy with minimal data transmission, adapts to dynamic environments, and effectively identifies uncertain samples. Code is provided in the supplementary materials.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6034