UniTAC: Universal Task-Aware Compression via Weighted Distortion Measures
A single learned codec can shift image fidelity toward a changing downstream task at runtime without retraining.
UniTAC sends a low-overhead importance vector as side information, conditioning both encoder and decoder while keeping one fixed backbone. The paper frames the method around weighted rate-distortion, then implements it with a Vision Transformer codec using token-level conditioning. In a localized task at 0.034 bpp, the model reports 91.4% accuracy, close to a task-based codec at 93.3% and above universal codecs at 76.9%. ArXiv · AI/CL/LG's note
UniTAC sends a low-overhead importance vector as side information, conditioning both encoder and decoder while keeping one fixed backbone. The paper frames the method around weighted rate-distortion, then implements it with a Vision Transformer codec using token-level conditioning. In a localized task at 0.034 bpp, the model reports 91.4% accuracy, close to a task-based codec at 93.3% and above universal codecs at 76.9%. ArXiv · AI/CL/LG's note
score 4