Scalable Multi-object Identification for Video Object Segmentation

We present a new semi-supervised video object segmentation framework that can process multiple objects in a single network pass and has a dynamically scalable architecture for speed-accuracy trade-offs. State-of-the-art methods prefer to match and segment a single positive object and have to process objects one by one under multi-object scenarios, consuming multiple times of computation resources. Besides, previous methods always have static network architectures, which are not flexible enough to adapt to different speed-accuracy requirements. To solve the above problems, we proposed an Associating Objects with Scalable Transformers (AOST) approach to match and segment multiple objects collaboratively with online network scalability. To match and segment multiple objects as efficiently as processing a single one, AOST employs an IDentification (ID) mechanism to assign objects with unique identities and associate them in a shared high-dimensional embedding space. In addition, a Scalable Long Short-Term Transformer (S-LSTT) is designed to construct hierarchical multi-object associations and enable online adaptation of accuracy-efficiency trade-offs. By further introducing scalable supervision and layer-wise ID-based attention, AOST is not only more flexible but more robust than previous methods. We conduct extensive experiments on multi-object and single-object benchmarks to evaluate AOST variants. Compared to state-of-the-art competitors, our methods can maintain superior run-time efficiency with better performance. Notably, we achieve new state-of-the-art performance on popular VOS benchmarks, i.e., YouTube-VOS (86.5%), DAVIS 2017 Val/Test (87.0%/84.7%), and DAVIS 2016 (93.0%). Project page: https://github.com/z-x-yang/AOT.