In short, the visual analysis further proves the necessity and effectiveness of our hierarchical LSTT. The hierarchical matching is not simply a combination of multiple matching processes.

In the semi-supervised setting, the first mask of each object is provided at test time.

Furthermore, we will integrate all changes according to your suggestions and questions. Figure 1: Evaluation on the DA VIS 2017 validation set. The connection between the inner and outer optimization is also illustrated in Algorithm 1 of the supplementary. The mitigate in line 7 refers to shortcomings and we will improve the understandability of the abstract. The fine-tuning epochs in Table 1 refer to a single update with one image. How to train your MAML.

Transformer is designed for constructing hierarchical matching and propagation. We conduct extensive experiments on both multi-object and single-object benchmarks to examine AOT variant networks with different complexities.

The optimization strategies and related hyper-parameters are also the same as AOT. The loss function is a 0.5:0.5 combination of BCE loss [ Such a process is necessary to keep enough long-term information and avoid facing out of memory when inferring long videos. The longest video in VOT 2020 contains 1,500 frames. We compare our DeAOT with more VOS methods in Table 2 and 1. VOS cases, including similar objects, occlusion, fast motion, motion blur, etc. A.4 Border Impact and Limitations The proposed DeAOT framework significantly improves VOS's performance, robustness, and robustness. As to limitations, the scenarios with multiple similar objects and severe occlusions are still very challenging for DeAOT and other VOS solutions.