Reproduce “SPA-Q: Structure-Preserving Adaptive Post-Training Quantization for Monocular Depth Estimation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jaemin Choi, Jincheol Yang, Nahyun Lim, Yun-Seong Jeong, Matti Zinke, Hyunwoo Yu, Suk-Ju Kang

SPA-Q targets low-bit quantization of monocular depth models for resource-constrained devices. It calibrates query-key attention distributions and learns channel-wise feature corrections, which are folded into weights after optimization. For Depth Anything v1 ViT-B on NYUv2, it improves 4-bit AbsRel from 0.1875 to 0.1021 and δ₁ from 0.6845 to 0.8991 versus QSCA. Quantized maps remain less clean than full-precision maps.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.