Boundary Perception Guidance: A Scribble-Supervised Semantic Segmentation Approach

Authors: Bin Wang, Guojun Qi, Sheng Tang, Tianzhu Zhang, Yunchao Wei, Linghui Li, Yongdong Zhang

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on the PASCAL VOC 2012 demonstrate the proposed BPG achieves m Io U of 73.2% without fully connected Conditional Random Field (CRF) and 76.0% with CRF, setting up the new state-of-the-art in literature.
Researcher Affiliation Collaboration Bin Wang1,3 , Guojun Qi2 , Sheng Tang1 , Tianzhu Zhang4 , Yunchao Wei5 , Linghui Li1,3 and Yongdong Zhang1 1Key Lab of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, China 2Huawei Cloud, China 3University of the Chinese Academy of Sciences, China 4University of Science and Technology of China, China 5University of Technology Sydney, Australia
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statement about releasing source code or a link to a code repository.
Open Datasets Yes We evaluate the proposed architecture on the PASCAL VOC 2012 segmentation benchmark [Everingham et al., 2010]... The training data are scribbles from [Lin et al., 2016] for the weak supervision task.
Dataset Splits Yes The original dataset contains 1, 461 training images, as well as 1, 449 validation and 1, 456 testing examples, respectively. Following the evaluation protocol in the literature [Chen et al., 2018a; Dai et al., 2015; Lin et al., 2016; Tang et al., 2018a], we use the augmented dataset by the extra annotations provided by [Hariharan et al., 2011], totaling 10, 582 training images.
Hardware Specification Yes We run 25 training epochs on a single NVIDIA Titan X 1080ti GPU, which takes about 10 hours on the PASCAL VOC dataset.
Software Dependencies No The paper mentions 'Py Torch' but does not provide a specific version number for it or any other software dependencies.
Experiment Setup Yes The proposed weakly-supervised semantic segmentation network is simply trained on a single scale of input images. Like the setting in deeplab-v2, we employ the poly learning rate policy for with a mini-batch of 10 images with an initial learning rate of 0.00025. We use a momentum of 0.9 and a weight decay of 0.0005. The hyper-parameter λ in Eq. 4 is set to 1.0.