ScholarWorks@성균관대학교

초록

Adapting deep learning models to neural processing units (NPUs) poses significant challenges due to limited operator support and hardware-specific constraints. In this paper, we present the deployment of two representative mod-els-PVN3D for object 6D pose estimation and HGGD for 6-DoF grasp detection-on Mobilint's MLA100 Low Profile equipped with the Aries2 NPU. To ensure compatibility, we replace unsupported operations and simplify architectural components while preserving task-specific performance. With these modifications, our NPU-deployed models retain over 96 % of their original accuracy while achieving significant power savings. We further analyze the impact of calibration dataset composition on quantization performance. Our results show that the relevance of calibration samples has a greater effect on accuracy than their quantity, highlighting the importance of using calibration data that are semantically consistent with the target domain. These findings offer practical insights into adapting complex models for deployment under NPU constraints.

키워드