Visual Boundary Knowledge Translation for Foreground Segmentation

Authors: Zunlei Feng, Lechao Cheng, Xinchao Wang, Xiang Wang, Ya Jie Liu, Xiangtong Du, Mingli Song1334-1342

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Exhaustive experiments demonstrate that, with only tens of labeled samples as guidance, Trans-Net achieves close results on par with fully supervised methods.
Researcher Affiliation Collaboration 1Zhejiang University 2Zhejiang Lab 3Stevens Institute of Technology 4Jiangsu University of Science and Technology {zunleifeng, xiang.wang, brooksong}@zju.edu.cn, {chenglc,liuyj}@zhejianglab.com, xinchao.wang@stevens.edu
Pseudocode Yes Algorithm 1 The Training Algorithm for Trans-Net
Open Source Code No The paper lists external datasets and references them, but does not contain an explicit statement about releasing its own source code or provide a link to a repository for the described methodology.
Open Datasets Yes The datasets we adopted contain single category datasets: Birds (Catherine et al. 2011), Flowers (Nilsback and Zisserman 2007) and Human Matting1, and mixed category datasets: THUR15K (Cheng et al. 2014), MSRA10K and MSRA-B (Cheng et al. 2011; Hou et al. 2017), CSSD (Yan et al. 2013), ECSSD (Shi et al. 2016), DUT-OMRON (Ruan, Tong, and Lu 2011), PASCALContext (Mottaghi et al. 2014), HKU-IS (Li and Yu 2016), SOD (Movahedi and Elder 2010), SIP1K (Fan et al. 2019).
Dataset Splits No The paper lists total sample counts for datasets (e.g., 'The Birds, Flowers, and Human Matting contain (11, 788), (8, 189), (34, 427) samples, respectively. The THUR15K contains 5 categories and 15000 samples.'), but does not explicitly provide specific percentages, sample counts, or predefined splits for training, validation, and test sets.
Hardware Specification No The paper specifies network architectures and parameter settings but does not explicitly describe the hardware (e.g., CPU, GPU models, or memory) used for running the experiments.
Software Dependencies No The paper mentions software components such as 'Deeplab V3+ (backbone: resnet50)' and 'Adam hyperparameters', but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Parameter setting. The parameters are set as follows: ξ = 1, τ = 1, η = 1, λ = 10, ncritic = 5, the batch size K = 64, Adam hyperparameters for two discriminators α = 0.0001, β1 = 0, β2 = 0.9. The learning rate for the segmentation network and two discriminators are all taken to be 1e 4. The disk strel of radius r is randomly sampled integer between 11 and 55.