Research map / Architectures and efficiency / Quantization
SPA-Q: Structure-Preserving Adaptive Post-Training Quantization for Monocular Depth Estimation
Jaemin Choi, Jincheol Yang, Nahyun Lim, Yun-Seong Jeong, Matti Zinke, Hyunwoo Yu, Suk-Ju Kang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SPA-Q, a structure-preserving adaptive post-training quantization method for monocular depth estimation foundation models such as Depth Anything.
Problem
Transformer-based monocular depth estimation foundation models (e.g., Depth Anything) are costly in compute and memory, and standard post-training quantization degrades them; the abstract is truncated, so the specific failure mode is only partially stated.
Approach
Quantization method named for preserving structure during adaptive post-training quantization of MDE models; the abstract is truncated so the exact mechanism is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| calibrate Q and K on pre-softmax logits instead of local per-projection error (ablation) | SPA-Q · NYUv2 W4A4, DA v1 ViT-B | AbsRel: 0.1908 -> 0.1839 | No clear effect |
| replace logits-level objective with attention-preserving calibration (APC) (ablation) | SPA-Q · NYUv2 W4A4, DA v1 ViT-B | AbsRel: 0.1839 -> 0.1444 | Improved |
| add channel-wise distribution alignment (CWDA) on top of APC (ablation) | SPA-Q · NYUv2 W4A4, DA v1 ViT-B | AbsRel: 0.1444 -> 0.1021 | Improved |
| add APC and CWDA vs local calibration (ablation) | SPA-Q · KITTI W4A4, DA v1 ViT-B | AbsRel: 0.3241 -> 0.1840 | Improved |
| use SPA-Q instead of QSCA under 4-bit | Depth Anything v1 ViT-B · NYUv2 | delta1: 0.6845 -> 0.8991 | Improved |
| use SPA-Q instead of QSCA under 4-bit | Depth Anything v1 ViT-B · KITTI | AbsRel: 0.2365 -> 0.1840 | Improved |
| use SPA-Q instead of QSCA on synthetic Sintel | Depth Anything v1 ViT-B · Sintel | AbsRel: 0.4781 -> 0.3949 | Improved |
| fuse CWDA affine parameters into weights after training (ablation) | SPA-Q · ViT-B efficiency | model size and BOPs: BOPs 4493 -> 4270 G | Improved |
| use SPA-Q instead of QSCA's auxiliary modules | Depth Anything v1 ViT-S · efficiency | model size: 16.27 MB -> 14.16 MB | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- QDrop — QDrop: Randomly Dropping Quantization for Extremely Low-bit Post-Training Quantization (ICLR 2022) · better on 2 of 2 shared results
- BRECQ — BRECQ: Pushing the Limit of Post-Training Quantization by Block Reconstruction (ICLR 2021) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Not All Tasks Quantize Equally: Fisher-Guided Quantization for Visual Geometry Transformer (NeurIPS 2026 · similarity 0.89)
- VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models (CVPR 2026 · similarity 0.88)
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · similarity 0.88)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.88)
- APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.87)
- MBQ: Modality-Balanced Quantization for Large Vision-Language Models (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
