Research map / Architectures and efficiency / Quantization
Transform-Free Feature Coding via Entropy-Constrained Vector Quantization
Qiaoxi Chen, Changsheng Gao, Li Li, Dong Liu
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 |
|---|---|---|---|
| remove the transform module and directly vector-quantize backbone features | feature coding pipeline · ResNet50-Cls, DINOv2-Cls, DINOv2-Seg | rate-distortion | Improved |
| replace Hyperprior transform codec with transform-free ECVQ | feature coding · ResNet50-Cls | MACs: 2.8G -> 0.4M | Improved |
| replace Hyperprior with transform-free ECVQ | feature coding · ResNet50-Cls | encoding time (ms): 16.1 -> 1.5 | Improved |
| replace VTM with transform-free ECVQ | feature coding · DINOv2 features, high bitrates | rate-distortion | Mixed |
| add entropy model and entropy constraint (ablation) | transform-free ECVQ · ResNet50-Cls | Top-1 accuracy: 74.29 -> 74.42 at BPFP 0.02 | No clear effect |
| add entropy model and entropy constraint (ablation) | transform-free ECVQ · DINOv2-Seg | mIoU: 68.37 (BPFP 0.39) w/o -> 52.13 (BPFP 0.02) with, wider rate range | Improved |
| increase chunk size at low bitrate (ablation) | transform-free ECVQ · DINOv2-Seg | mIoU | Improved |
| increase chunk size at high bitrate (ablation) | transform-free ECVQ · DINOv2-Seg | mIoU | Hurt |
| increase codebook size for ResNet50 features (ablation) | transform-free ECVQ · ResNet50-Cls | codebook utilization: 100.0% -> 80.9% | Hurt |
Closest papers by meaning
- Differentiable Vector Quantization for Rate-Distortion Optimization of Generative Image Compression (CVPR 2026 · similarity 0.85)
- NVTC: Nonlinear Vector Transform Coding (CVPR 2023 · similarity 0.85)
- Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding (ICLR 2025 · similarity 0.85)
- HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression (NeurIPS 2026 · similarity 0.84)
- Neural Video Compression with Feature Modulation (CVPR 2024 · similarity 0.84)
- Linear Attention Modeling for Learned Image Compression (CVPR 2025 · 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.
