An Effective Entropy Model for Semantic Feature Compression

Semantic feature compression aims to compress image features for downstream machine vision tasks without reconstructing image pixels. Such a task is very challenging since it needs to learn features which are not only useful for machine vision tasks, but also easy to compress. While existing learnable feature coding models utilize downstream task networks as teacher networks to guide the learning and compression of semantic features, they use simple entropy models and do not effectively reduce information redundancy. In this work, we propose a transformer-based spatial-channel auto-regressive feature context model (SC-AR FCM) to assist the entropy coding of learnable features. Through extensive experimentation on object detection and segmentation tasks, we demonstrate that the rate-accuracy performance of our proposed method surpasses traditional image compression techniques and state-of-the-art learning-based feature compression techniques.