Decoupling Features in Hierarchical Propagation for Video Object Segmentation
Authors: Zongxin Yang, Yi Yang
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that De AOT significantly outperforms AOT in both accuracy and efficiency. On You Tube-VOS, De AOT can achieve 86.0% at 22.4fps and 82.0% at 53.4fps. Without test-time augmentations, we achieve new state-of-the-art performance on four benchmarks, i.e., You Tube VOS (86.2%), DAVIS 2017 (86.2%), DAVIS 2016 (92.9%), and VOT 2020 (0.622). |
| Researcher Affiliation | Collaboration | Zongxin Yang1,2, Yi Yang1 1 CCAI, College of Computer Science and Technology, Zhejiang University 2 Baidu Research {yangzongxin, yangyics}@zju.edu.cn |
| Pseudocode | No | The paper describes the methodology using mathematical formulations and architectural diagrams (e.g., Figure 3), but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Project page: https://github.com/z-x-yang/AOT. |
| Open Datasets | Yes | We conduct experiments on three popular VOS benchmarks (You Tube-VOS [57], DAVIS 2017 [39], and DAVIS 2016 [38]) and one challenging Visual Object Tracking (VOT) benchmark (VOT 2020 [24]). |
| Dataset Splits | Yes | You Tube-VOS [57] is a large-scale multi-object VOS benchmark, which contains 3471 videos in the training split with 65 categories and 474/507 videos in the Validation 2018/2019 split with additional 26 unseen categories. |
| Hardware Specification | Yes | All the results were fairly recorded on the same device, 1 Tesla V100 GPU. |
| Software Dependencies | No | The paper mentions encoders (Mobile Net-V2, Res Net-50, Swin-B) and a decoder (FPN) and channel dimensions, but does not specify software dependencies with version numbers such as PyTorch, TensorFlow, or CUDA versions within the main text provided. |
| Experiment Setup | Yes | In our GPM module, the channel dimension C of visual and ID embeddings is 256, the matching features dimension Ck is 128, and the propagation features dimension Cv is 512. Moreover, the kernel size of Fdw is 5, and the gating function σ( ) is Si LU/Swish [18,41]. To make fair comparisons with AOT s variants [63], we build corresponding De AOT variants with different GPM number L or long-term memory size m. The hyper-parameters of these variants are: De AOT-T: L = 1, m = {1}; De AOT-S: L = 2, m = {1}; De AOT-B: L = 3, m = {1}; De AOT-L: L = 3, m = {1, 1+δ, 1+2δ, ...}. |