Research map / Detection and segmentation / 3D detection and driving
FQ-PETR: Fully Quantized Position Embedding Transformation for Multi-View 3D Object Detection
Jiangyong Yu, Changyong Shu, Sifan Zhou, Zichen Yu, Xing Hu, Dawei Yang
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 |
|---|---|---|---|
| replace camera-ray PE with QFPE | PETR (INT8 QuaRot) · nuScenes val | mAP: 22.81 -> 31.46 | Improved |
| replace camera-ray PE with QFPE in FP32 (ablation) | PETR · nuScenes val | NDS: 36.11 -> 37.20 | Improved |
| replace camera-ray PE with QFPE in FP32 (ablation) | StreamPETR · nuScenes val | mAP: 49.51 -> 50.48 | Improved |
| use 3 anchor embeddings per axis instead of 2, 4 or 5 (ablation) | QFPE · nuScenes val | NDS: 36.66 -> 37.20 | Improved |
| quantize softmax after numerical stabilization (N=20) (ablation) | PETR · nuScenes val | mAP: 13.79 -> 31.42 | Improved |
| increase truncation range N from 10 to 20 (ablation) | QANS · nuScenes val | mAP: 29.33 -> 31.42 | Improved |
| increase truncation range N beyond 20 (ablation) | QANS · nuScenes val | mAP: 31.42 -> 31.42 | No clear effect |
| use DULUT instead of linear LUT with 128 entries (ablation) | PETR · nuScenes val | mAP: 31.05 -> 31.42 | Improved |
| use DULUT (32,32) vs (16,16) entries (ablation) | DULUT · nuScenes val | mAP: 17.36 -> 31.36 | Improved |
| use DULUT instead of I-BERT / I-ViT (ablation) | PETR · nuScenes val | mAP: 28.77 -> 31.36 | Improved |
| do not quantize softmax input (keep FP) (ablation) | PETR camera-ray PE, SmoothQuant INT8 · nuScenes val | mAP: 20.67 -> 24.90 | Improved |
| use QFPE PTQ instead of distillation QAT (ablation) | PETR · nuScenes val | mAP: 30.3 -> 31.46 | Improved |
Closest papers by meaning
- LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection (ICLR 2024 · similarity 0.87)
- Not All Tasks Quantize Equally: Fisher-Guided Quantization for Visual Geometry Transformer (NeurIPS 2026 · similarity 0.87)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.85)
- APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.85)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.85)
- Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector (CVPR 2024 · similarity 0.84)
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.
