Compressing Multi-Scale Features with a Channel-Shrinked Single-Branch Architecture

Neural network feature compression is an important technology in the field of collaborative intelligence. In recent years, various feature compression technologies have been developed to target a diverse range of tasks and neural networks. Since neural networks equipped with Feature Pyramid Networks (FPN) usually performed better in certain scenarios, many feature coding technologies are designed specifically for such networks, where multiple feature tensors with different scales are encoded and reconstructed. Current feature coders typically adopt a multi-branch paradigm with each branch responsible for processing one FPN tensor. However, designing multi-branch neural networks is difficult, since determining the optimal amount of weight allocated to each feature tensor or shared among multiple tensors requires a substantial amount of experiments, which is unaffordable to most researchers. Consequently, existing multi-branch feature coding models only achieved sub-optimal performances. To tackle the above problem, we propose to encode multi-scale features with one single-branch architecture. Such architecture was formed semi-automatically with an adaptive channel shrinking mechanism to spare the extensive experiments that are typically required. The proposed architecture achieved 89.63% and 90.34% BD rate gains compared to H.266/VVC on object detection and instance segmentation respectively. Our codes are released at: https://github.com/pusheng0902/singleBranchFeatureCoding.