Research map / Detection and segmentation / CNN features
Fully Quantized Network for Object Detection
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 batch norm freezing throughout quantization-aware fine-tuning (ablation) | 4-bit ResNet-18 RetinaNet · COCO | AP: 0.197 -> 0.235 | Helped |
| Use percentile activation calibration instead of moving-average statistics (ablation) | 4-bit ResNet-18 RetinaNet · COCO | AP: 0.197 -> 0.222 | Helped |
| Use channel-wise weight quantization instead of layer-wise quantization (ablation) | 4-bit ResNet-18 RetinaNet · COCO | AP: 0.197 -> 0.250 | Helped |
| Add batch norm freezing, percentile activation calibration, and channel-wise quantization (ablation) | 4-bit ResNet-18 RetinaNet · COCO | AP: 0.197 -> 0.286 | Helped |
| Add batch norm freezing (ablation) | 8-bit ResNet-18 RetinaNet · COCO | AP: 0.296 -> 0.299 | Helped |
| Use percentile activation calibration (ablation) | 8-bit ResNet-18 RetinaNet · COCO | AP: 0.296 -> 0.313 | Helped |
| Use channel-wise weight quantization (ablation) | 8-bit ResNet-18 RetinaNet · COCO | AP: 0.296 -> 0.293 | Hurt |
| Freeze batch norm statistics for the entire fine-tuning process rather than only the final 10k steps (ablation) | ResNet-18 RetinaNet · COCO | AP: 0.226 -> 0.236 | Helped |
| Vary the activation-calibration percentile parameter (ablation) | 4-bit ResNet-18 RetinaNet · COCO | AP: 0.286 -> 0.289; 0.286 -> 0.275 | Mixed |
| Quantize the detector to 4-bit with FQN instead of full precision | ResNet-50 RetinaNet · COCO-2017-val | mAP: 0.356 -> 0.325 | Hurt |
| Quantize the detector to 4-bit with FQN instead of full precision | ResNet-50 Faster R-CNN · COCO-2017-val | mAP: 0.377 -> 0.331 | Hurt |
| Replace integer-only method [17] with FQN | 4-bit ResNet-18 RetinaNet · COCO | AP: 0.197 -> 0.286 | Helped |
Closest papers by meaning
- AQD: Towards Accurate Quantized Object Detection (CVPR 2021 · similarity 0.91)
- Bit-Shrinking: Limiting Instantaneous Sharpness for Improving Post-Training Quantization (CVPR 2023 · similarity 0.88)
- Quantization Networks (CVPR 2019 · similarity 0.88)
- Post training 4-bit quantization of convolutional networks for rapid-deployment (NeurIPS 2019 · similarity 0.88)
- Towards Efficient Verification of Quantized Neural Networks (AAAI 2024 · similarity 0.87)
- Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector (CVPR 2024 · 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.