CBNetV2: A Composite Backbone Network Architecture for Object Detection

Modern top-performing object detectors depend heavily on backbone networks, whose advances bring consistent performance gains through exploring more effective network structures. However, designing or searching for a new backbone and pre-training it on ImageNet may require a large number of computational resources, making it costly to obtain better detection performance. In this paper, we propose a novel backbone network, namely CBNetV2, by constructing compositions of existing open-sourced pre-trained backbones. In particular, CBNetV2 architecture groups multiple identical backbones, which are connected through composite connections. We also propose a better training strategy with the Assistant Supervision for CBNet-based detectors. Without additional pre-training, CBNetV2 can be integrated into mainstream detectors, including one-stage and two-stage detectors, as well as anchor-based and anchor-free-based ones, and significantly improve their performance by more than 3.0% AP over the baseline on COCO. Also, experiments provide strong evidence showing that composite backbones are more efficient and resource-friendly than pre-trained wider and deeper networks, including manual-based and NAS-based, as well as CNN-based and Transformer-based ones. Particularly, with single-model and single-scale testing, our HTC Dual-Swin-B achieves 58.6% box AP and 51.1% mask AP on COCO test-dev, which is significantly better than the state-of-the-art result (i.e., 57.7% box AP and 50.2% mask AP) achieved by a stronger baseline HTC++ with a larger backbone Swin-L. Code will be released at https://github.com/VDIGPKU/CBNetV2.