Research map / Architectures and efficiency / Pruning and efficient tuning
TransTIC: Transferring Transformer-based Image Compression from Human Perception to Machine Perception
Yi-Hsin Chen, Ying-Chieh Weng, Chia-Hao Kao, Cheng Chien, Wei-Chen Chiu, Wen-Hsiao Peng
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 TransTIC prompting instead of spatial feature transform on complex tasks (detection) | TIC base codec · COCO2017-val detection | BD-Rate (%): -26.56 -> -46.07 | Improved |
| use TransTIC prompting instead of SFT on classification | TIC base codec · ImageNet-val classification | BD-Rate (%): -62.10 -> -58.77 | No clear effect |
| use prompting instead of channel selection | TIC base codec · COCO2017-val detection | BD-Rate (%): 53.66 -> -46.07 | Improved |
| use full fine-tuning instead of prompting | TIC base codec · ImageNet-val classification | BD-accuracy: 10.02 -> 16.24 | Improved |
| swap recognition network to VGG19 at test time (full fine-tuning vs TransTIC) | full fine-tuning codec · ImageNet-val classification | accuracy: full fine-tuning drops 15%-20%, TransTIC drops 3%-7% | Hurt |
| instance-specific (IP-type) instead of task-specific (TP-type) prompts in encoder (ablation) | TransTIC encoder · classification and detection | rate-accuracy | Improved |
| add decoder-side task-specific prompts (ablation) | TransTIC · classification and detection | rate-accuracy | Improved |
| use spatially adaptive instead of shared instance prompts across windows (ablation) | TransTIC encoder · classification and detection | rate-accuracy | Improved |
| inject prompts to early STBs (1,2) rather than later STBs (3,4) (ablation) | TransTIC encoder · classification and detection | rate-accuracy | Improved |
| inject prompts into all four STBs instead of STB-1,2 (ablation) | TransTIC encoder · classification and detection | rate-accuracy | No clear effect |
| increase prompts per window from 16 to 64 (ablation) | TransTIC encoder · classification and detection | accuracy | No clear effect |
| use ROI coding instead of learned spatially adaptive transfer | image codec for machines · ImageNet-val classification | BD-Rate (%): -58.77 -> -8.67 | Hurt |
Later papers that compared with this method
1 later accepted paper reports TransTIC as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Image Compression for Machine and Human Vision With Spatial-Frequency Adaptation (ECCV 2024 · 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 0.85)
- Controllable Distortion-Perception Tradeoff Through Latent Diffusion for Neural Image Compression (AAAI 2025 · similarity 0.84)
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.84)
- Correcting Diffusion-Based Perceptual Image Compression with Privileged End-to-End Decoder (ICML 2024 · similarity 0.83)
- Unified Coding for Both Human Perception and Generalized Machine Analytics with CLIP Supervision (AAAI 2025 · similarity 0.83)
- Towards Unified Human Perception and Machine Understanding: Token Flow Guided Compression Framework (CVPR 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.
