Research map / Architectures and efficiency / Quantization
Video Coding for Machines: Compact Visual Representation Compression for Intelligent Collaborative Analytics
Wenhan Yang, Haofeng Huang, Yueyu Hu, Ling‐Yu Duan, Jiaying 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 |
|---|---|---|---|
| use codebook hyperprior instead of IDFC | feature compression · Semantic segmentation | bit-rate (bpp): 0.026 -> 0.013 | Improved |
| use codebook hyperprior instead of plain hyperprior | feature compression · Semantic segmentation | mIoU: 25.42% -> 29.35% | Improved |
| use codebook hyperprior instead of plain hyperprior | feature compression · Scene classification | bit-rate (bpp): 0.0102 -> 0.0069 | Improved |
| use codebook hyperprior instead of IDFC | feature compression · Scene classification | test accuracy: 65.43% -> 59.47% | Hurt |
| use codebook hyperprior instead of IDFC | feature compression · Reshading | bit-rate (bpp): 0.0959 -> 0.0134 | Improved |
| compress tasks jointly in grouped Trinity setting instead of per-task Customized (ablation) | multi-task compression · Six Taskonomy tasks | total bit-rate (bpp): 0.059 -> 0.049 | Improved |
| group six tasks in Trinity instead of non-optimal Trinity* (ablation) | multi-task compression · Scene classification | accuracy: 59.64% -> 71.08% | Improved |
| compress all six tasks in one model (Hex) instead of Trinity (ablation) | multi-task compression · Scene classification | accuracy: 71.08% -> 62.18% | Hurt |
| compress all six tasks in one model (Hex) instead of Trinity (ablation) | multi-task compression · Semantic segmentation | mIoU: 18.14% -> 20.30% | Improved |
| add full pixel reconstruction to the Trinity group (Trinity+) (ablation) | multi-task compression · Scene classification | accuracy and bpp: 71.08% -> 64.62%; 0.016 -> 0.052 | Hurt |
| use BPG image codec instead of learned task features | BPG-43 · Scene classification | accuracy: 71.08% -> 50.05% | Hurt |
| add source feature of unseen task to the compression (Source+) vs Internal (ablation) | unseen task support · Object classification | accuracy: 51.06% -> 53.50% | Improved |
Closest papers by meaning
- VNVC: A Versatile Neural Video Coding Framework for Efficient Human-Machine Vision (TPAMI 2024 · similarity 0.90)
- All-in-One Image Coding for Joint Human-Machine Vision with Multi-Path Aggregation (NeurIPS 2024 · similarity 0.89)
- Unified Coding for Both Human Perception and Generalized Machine Analytics with CLIP Supervision (AAAI 2025 · similarity 0.89)
- PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models (CVPR 2025 · similarity 0.88)
- Benchmarking and Enhancing VLM for Compressed Image Understanding (ICML 2026 · similarity 0.87)
- When MLLMs Meets Compression Distortion: A Coding Paradigm Tailored to MLLMs (ICLR 2026 · 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.
