Reliability Does Matter: An End-to-End Weakly Supervised Semantic Segmentation Approach

Authors: Bingfeng Zhang, Jimin Xiao, Yunchao Wei, Mingjie Sun, Kaizhu Huang12765-12772

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Despite its apparent simplicity, our one-step solution achieves competitive m Io U scores (val: 62.6, test: 62.9) on Pascal VOC compared with those two-step state-of-the-arts. By extending our one-step method to two-step, we get a new state-of-the-art performance on the Pascal VOC (val: 66.3, test: 66.5). Experiments Dataset and Implementation Details Dataset. Our RRM is trained and validated on PASCAL VOC 2012 (Everingham et al. 2010) as well as its augmented data, including 10, 582 images for training, 1, 449 images for validating and 1, 456 images for testing. Mean intersection over union (m Io U) is considered as the evaluation criterion.
Researcher Affiliation Collaboration 1Xi an Jiaotong-liverpool University, Suzhou, China 2University of Technology Sydney, Australia 3Alibaba-Zhejiang University Joint Institute of Frontier Technologies, Hangzhou, China 1{Bingfeng.Zhang, jimin.xiao, mingjie.sun18, kaizhu huang}@xjtlu.edu.cn, 2yunchao.wei@uts.edu.au
Pseudocode Yes The algorithm flow is illustrated in Algorithm 1. Algorithm 1 Algorithm flow of our proposed approach.
Open Source Code Yes Code now is available at: https://github.com/zbf1991/RRM.
Open Datasets Yes Our RRM is trained and validated on PASCAL VOC 2012 (Everingham et al. 2010)
Dataset Splits Yes Our RRM is trained and validated on PASCAL VOC 2012 (Everingham et al. 2010) as well as its augmented data, including 10, 582 images for training, 1, 449 images for validating and 1, 456 images for testing.
Hardware Specification Yes All the experiments were performed on NVIDIA RTX 2080 Ti.
Software Dependencies No The paper states: 'Reproducibility: Py Torch (Paszke et al. 2017) was used.' While PyTorch is mentioned, a specific version number for the software is not provided. No other software dependencies with version numbers are listed.
Experiment Setup Yes The training learning rate is 0.001 with weight decay being 5e-4. The training images are resized with a ratio randomly sampled from (0.7, 1.3), and they are randomly flipped. Finally, they are normalized and randomly cropped to size 321*321. To generate reliable regions, the scale ratio in (2) is set to {0.5, 1, 1.5, 2}, γ in (5) is set to 4 for Pfg bg. The CRF parameters in (5) follow the setting in (Ahn and Kwak 2018). In (6), an α value is chosen with 40% pixels selected as labeled pixels for each class. During validating and testing, dense CRF is applied as a post-processing method, and the parameters are set as the default values given in (Huang et al. 2018). σd and σr in our dense energy loss are set as 15 and 100, respectively.