Research map / Architectures and efficiency / Pruning and efficient tuning
Image Compression for Machine and Human Vision With Spatial-Frequency Adaptation
Han Li, Shaohui Li, Shuangrui Ding, Wenrui Dai, Maida Cao, Chenglin Li, Junni Zou, Hongkai Xiong
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 |
|---|---|---|---|
| add the frequency modulation adapter alongside the spatial modulation adapter (ablation) | SFMA on mbt2018-mean · ImageNet-val classification | BD-acc: 16.30 -> 18.71 | Improved |
| add the spatial modulation adapter alongside the frequency modulation adapter (ablation) | SFMA on mbt2018-mean · ImageNet-val classification | BD-acc: 17.29 -> 18.71 | Improved |
| use FMA in parallel with SMA instead of a second parallel SMA (ablation) | SFMA on mbt2018-mean · ImageNet-val classification | BD-acc: 17.59 -> 18.71 | Improved |
| parallel rather than sequential arrangement of FMA and SMA (ablation) | SFMA on mbt2018-mean · Classification, detection, segmentation | BD-rate/BD-acc: classification BD-acc 18.67 and 18.50 (sequential) vs 18.71 | Mixed |
| place SFMA at shallow stage 1 rather than stage 3 (ablation) | SFMA on mbt2018-mean · ImageNet-val classification | BD-rate: -64.76% -> -69.54% | Improved |
| plug SFMA into all three stages rather than one (ablation) | SFMA on mbt2018-mean · ImageNet-val classification | BD-rate: -69.54% -> -82.00% | Improved |
| add SFMA to the decoder as well as the encoder (ablation) | SFMA on mbt2018-mean · ImageNet-val classification | BD-acc: 13.28 -> 18.71 | Improved |
| add SFMA to the encoder as well as the decoder (ablation) | SFMA on mbt2018-mean · ImageNet-val classification | BD-acc: 12.40 -> 18.71 | Improved |
| increase middle dimension from 64 to 128 (ablation) | SFMA on mbt2018-mean · Detection COCO2017-val | BD-mAP: 3.84 -> 4.11 (params 0.28M -> 0.62M) | Improved |
| increase middle dimension from 1 to 64 (ablation) | SFMA on mbt2018-mean · ImageNet-val classification | BD-acc: 12.58 -> 18.71 | Improved |
| use SFMA adapters instead of full fine-tuning | Lu2022-TIC base codec · Classification, detection, segmentation | BD-acc / BD-mAP: classification 17.68 -> 16.90; detection 4.51 -> 3.55 with 0.28M vs 7.51M params | Mixed |
| use SFMA instead of channel selection | Lu2022-TIC base codec · Detection COCO2017-val | BD-mAP: -0.55 -> 3.55 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TransTIC — TransTIC: Transferring Transformer-based Image Compression from Human Perception to Machine Perception (ICCV 2023) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability (CVPR 2025 · similarity 0.86)
- Unified Coding for Both Human Perception and Generalized Machine Analytics with CLIP Supervision (AAAI 2025 · similarity 0.84)
- Frequency-Aware Transformer for Learned Image Compression (ICLR 2024 · similarity 0.83)
- Frequency-Biased Synergistic Design for Image Compression and Compensation (CVPR 2025 · similarity 0.83)
- Frequency Dynamic Convolution for Dense Image Prediction (CVPR 2025 · similarity 0.82)
- F-Adapter: Frequency-Adaptive Parameter-Efficient Fine-Tuning in Scientific Machine Learning (NeurIPS 2025 · similarity 0.82)
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.
