Research map / Architectures and efficiency / Quantization
AHCPTQ: Accurate and Hardware-Compatible Post-Training Quantization for Segment Anything Model
Wenlun Zhang, Yunshan Zhong, Shimpei Ando, Kentaro Yoshioka
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CAG to AHCPTQ without it (HLUQ only) (ablation) | SAM-L · COCO instance segmentation, YOLOX prompts, W4A4 | mAP: 5.7 -> 31.0 | Improved |
| add CAG to AHCPTQ without it (HLUQ only) (ablation) | SAM-H · COCO instance segmentation, YOLOX prompts, W4A4 | mAP: 24.8 -> 35.2 | Improved |
| add HLUQ to AHCPTQ without it (CAG only) (ablation) | SAM-L · COCO instance segmentation, YOLOX prompts, W4A4 | mAP: 22.7 -> 31.0 | Improved |
| add HLUQ to AHCPTQ without it (CAG only) (ablation) | SAM-H · COCO instance segmentation, YOLOX prompts, W4A4 | mAP: 20.2 -> 35.2 | Improved |
| add CAG alone to baseline (ablation) | SAM-L · COCO instance segmentation, YOLOX prompts, W4A4 | mAP: 2.5 -> 22.7 | Improved |
| add HLUQ alone to baseline (ablation) | SAM-B · COCO instance segmentation, YOLOX prompts, W4A4 | mAP: 3.7 -> 5.7 | Improved |
| use HLUQ instead of biased log2 quantizer (ablation) | SAM-H · COCO instance segmentation, YOLOX prompts, W4A4 | mAP: 34.3 -> 35.2 | Improved |
| use HLUQ instead of uniform quantizer (ablation) | SAM-L · COCO instance segmentation, YOLOX prompts, W4A4 | mAP: 20.2 -> 31.0 | Improved |
| increase CAG group number from 1 (per-tensor) to 4 (ablation) | SAM-B · COCO YOLOX W4A4 | mAP: 5.7 -> 13.4 | Improved |
| use per-channel instead of 4 groups in CAG (ablation) | SAM-B · COCO YOLOX W4A4 | mAP: 13.4 -> 16.9 | Mixed |
| use AHCPTQ INT4 instead of FP32 accelerator | MLP block FPGA · FPGA MLP block of SAM mask decoder | speedup: 1.00× -> 7.89× | Improved |
| replace PTQ4SAM with AHCPTQ | SAM-L DINO · COCO DINO W4A4 | mAP: 2.3 -> 36.6 | 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 3 of 3 shared results
- BRECQ — BRECQ: Pushing the Limit of Post-Training Quantization by Block Reconstruction (ICLR 2021) · better on 3 of 3 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
- PTQ4SAM: Post-Training Quantization for Segment Anything (CVPR 2024 · similarity 0.92)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.88)
- TinySAM: Pushing the Envelope for Efficient Segment Anything Model (AAAI 2025 · similarity 0.88)
- Instance-Aware Group Quantization for Vision Transformers (CVPR 2024 · similarity 0.86)
- APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.86)
- Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector (CVPR 2024 · similarity 0.86)
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.
