Research map / Architectures and efficiency / Quantization
ORFC: Orthogonal Reparameterization for Low-Bitrate ViT Feature Coding
Letian Zhang, Wenhan Yang, Lingyu Duan
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 |
|---|---|---|---|
| learn an orthogonal rotation under output-deviation loss before grouped PQ (ablation) | ORFC · DINOv2 ViT-L/14 block 20 segmentation | BD-rate vs VTM (%): 14.4% -> -41.4% | Improved |
| learn an orthogonal rotation under output-deviation loss before grouped PQ (ablation) | ORFC · DINOv2 ViT-L/14 block 20 classification | BD-rate vs VTM (%): -57.0% -> -77.5% | Improved |
| use frozen-tail output deviation Lref instead of feature-space MSE as distortion (ablation) | ORFC · DINOv2 ViT-L/14 block 20 segmentation | BD-rate vs VTM (%): 153.2% -> -41.4% | Improved |
| use frozen-tail output deviation Lref instead of feature-space MSE as distortion (ablation) | ORFC · DINOv2 ViT-L/14 block 20 classification | BD-rate vs VTM (%): -5.2% -> -77.5% | Improved |
| add soft assignment with temperature annealing (ablation) | ORFC · DINOv2 ViT-L/14 block 20 segmentation | BD-rate vs VTM (%): 181.2% -> -41.4% | Improved |
| add soft assignment with temperature annealing (ablation) | ORFC · DINOv2 ViT-L/14 block 20 classification | BD-rate vs VTM (%): -24.8% -> -77.5% | Improved |
| ORFC vs OPQ at matched low bitrate | DINOv2 ViT-L/14 block 10 · ImageNet cls | Acc@1 (%): 34.2 -> 96.0 | Improved |
| ORFC vs OPQ and VTM at low bitrate | VitDet ViT-B/16 block 3 · COCO2017 detection | mAP@0.5:0.95 (%): 36.7 -> 91.6 (OPQ), 49.5 -> 91.6 (VTM) | Improved |
| ORFC vs VTM on deeper layers | DINOv2 ViT-L/14 · VOC2012 seg | BD-rate vs VTM (%): -65.8 (blk10) vs -41.4 (blk20) | Mixed |
| MSE-driven rotation (OPQ) instead of task-aligned rotation | grouped PQ · DINOv2 ViT-L/14 block 20 group-wise sensitivity | imbalance of per-group accuracy gain | No clear effect |
Closest papers by meaning
- Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient (CVPR 2025 · similarity 0.84)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.84)
- Neural Video Compression with Feature Modulation (CVPR 2024 · similarity 0.84)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.84)
- RePaViT: Scalable Vision Transformer Acceleration via Structural Reparameterization on Feedforward Network Layers (ICML 2025 · similarity 0.83)
- Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.83)
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.
