Research map / Detection and segmentation / CNN features
Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector
Read the paperCheck your idea against 108,967 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 Reg-PTQ, a post-training quantization framework that fully quantizes object detectors (including heads) using a filtered global-loss calibration and a learnable logarithmic-affine quantizer.
Problem
Post-training quantization of object detectors usually leaves the detection head unquantized, and classical PTQ methods degrade severely when the regression head is quantized at low bit-width.
Approach
Argues regressors are more sensitive to quantization noise and have non-uniform, Laplace/Gaussian-like weight distributions due to distance-based losses; combines local reconstruction loss with a global detection loss filtered by confidence and IoU thresholds (FGIC), and uses a learnable log-affine transform before uniform quantization (LLAQ).
Key results
- On COCO under W4A8, RetinaNet ResNet-101 reaches 38.6 mAP vs 38.9 full precision (only 0.3% drop); Faster RCNN drops only 0.7%.
- Under W2A4 on RetinaNet ResNet-50, Reg-PTQ achieves 23.9 mAP, up to 4.0 points above prior state-of-the-art PTQ methods.
- Full quantization of Faster RCNN ResNet-50 at W2A4 gives 14.2x FLOPs and 7.8x storage reduction vs full precision, with an extra 2.1x and 3.6x over quantizing only backbone and neck.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Reg-PTQ instead of the strongest listed baseline | RetinaNet ResNet-50 · COCO W4A4 | mAP: 34.2 -> 36.7 | Helped |
| use Reg-PTQ instead of the strongest listed baseline | YOLOF ResNet-50 · COCO W4A4 | mAP: 33.4 -> 34.3 | Helped |
| use Reg-PTQ instead of the strongest listed baseline | Faster R-CNN ResNet-50 · COCO W4A4 | mAP: 33.3 -> 36.7 | Helped |
| use Reg-PTQ instead of the strongest listed baseline | Faster R-CNN ResNet-101 · COCO W4A4 | mAP: 35.4 -> 36.2 | Helped |
| use Reg-PTQ instead of the strongest listed baseline | Mask R-CNN ResNet-50 · COCO W4A4 | mAP: 34.9 -> 36.4 | Helped |
| use Reg-PTQ instead of the strongest listed baseline | Mask R-CNN ResNet-101 · COCO W4A4 | mAP: 36.8 -> 37.2 | Helped |
| add global loss to local loss without filtering (ablation) | FGIC on RetinaNet ResNet-50 · COCO W2A4 | mAP: 23.0 -> 23.5 | Helped |
| add two-step filtering of low-confidence and low-IoU boxes (ablation) | FGIC on RetinaNet ResNet-50 · COCO W2A4 | mAP: 23.5 -> 23.9 | Helped |
| replace uniform quantization with LLAQ (ablation) | RetinaNet ResNet-50 head · COCO W2A4 | mAP: 23.0 -> 23.6 | Helped |
| replace uniform quantization with LLAQ (ablation) | RetinaNet ResNet-50 head · COCO W3A3 | mAP: 27.2 -> 28.0 | Helped |
| replace uniform quantization with LLAQ (ablation) | RetinaNet ResNet-50 head · COCO W4A4 | mAP: 35.2 -> 35.7 | Helped |
| replace uniform quantization with LLAQ (ablation) | Faster R-CNN ResNet-50 heads · COCO W2A4 | mAP: 22.3 -> 23.7 | Helped |
Surprising findings
- Local quantization error (MSE) minimization does not select the best scaling factor for regressors, unlike classifiers.
- Diagonal Fisher approximation of the Hessian fails to track quantized regressor performance, while the precise Hessian does.
Limitations
- Relies on a hand-set confidence threshold and IoU threshold for filtering; sensitivity was only checked on RetinaNet R50 under W2A4.
- Efficiency is measured as FLOPs and storage via thop, not on real integer hardware latency; RetinaNet FLOPs reduction is not noticeable.
- Accuracy drops remain noticeable at W4A4 (about 3% on some detectors) and at very low bit-widths.
Experimental setup
Datasets: COCO PASCAL VOC Iris flower dataset
Benchmarks: COCO PASCAL VOC
Metrics: mAP FLOPs storage
Baselines: AdaRound AdaQuant BRECQ QDrop PD-Quant SubSetQ
Builds on: BRECQ QDrop AdaRound Hessian-guided metric calibration
Closest papers by meaning
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.89)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.87)
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · similarity 0.87)
- LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection (ICLR 2024 · similarity 0.87)
- ERQ: Error Reduction for Post-Training Quantization of Vision Transformers (ICML 2024 · similarity 0.86)
- PTQ4SAM: Post-Training Quantization for Segment Anything (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.