Renovate Yourself: Calibrating Feature Representation of Misclassified Pixels for Semantic Segmentation

Authors: Hualiang Wang, Huanpeng Chu, Siming FU, Zuozhu Liu, Haoji Hu2450-2458

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that our method could significantly boost the performance of current segmentation methods on multiple datasets (e.g., we outperform the original HRNet and OCRNet by 1.1% and 0.9% m Io U on the Cityscapes test set).
Researcher Affiliation Collaboration Hualiang Wang1*, Huanpeng Chu1*, Siming Fu1, Zuozhu Liu2, Haoji Hu1 1 College of Information Science and Electronic Engineering, Zhejiang University, China 2 Zhejiang University-Angelalign Inc. R&D Center for Intelligent Healthcare, ZJU-UIUC Institute, Zhejiang University, China. {hualiang wang,chuhp,fusiming,haoji hu}@zju.edu.cn, zuozhuliu@intl.zju.edu.cn
Pseudocode No No pseudocode or algorithm blocks are provided in the paper.
Open Source Code Yes Codes are available at https://github.com/Vipai Lab/RCH.
Open Datasets Yes We conduct experiments on Cityscapes (Cordts et al. 2016), ADE20K (Zhou et al. 2017) and Pascal Context (Mottaghi et al. 2014).
Dataset Splits Yes The Cityscapes contains 19 categories from 5000 images of high resolution (2048 1024), of which 2975 images for training, 500 images for validation and 1525 for testing. The ADE20K is a scene parsing dataset covering 150 classes from 20210 images. The dataset is divided into 20K/2K/3K images for training, validation and testing, respectively. The Pascal Context dataset contains 59 semantic classes and 1 background class. The training set and test set consist of 4998 and 5105 images respectively.
Hardware Specification Yes For reproducibility, we use mmsegmentation (Contributors 2020) as our codebase and the networks are trained with 8 Nvidia Titan XP.
Software Dependencies No The paper mentions using 'mmsegmentation (Contributors 2020)' as its codebase but does not provide specific version numbers for this or other software dependencies.
Experiment Setup Yes We train the models using Adam optimizer with the initial learning rate 0.01, weight decay 0.0005 and momentum 0.9. The learning rate dynamically decays exponentially according to the ploy strategy. To provide a fair comparison, we adopt the widely-used tricks: OHEM (Shrivastava, Gupta, and Girshick 2016) and auxiliary loss (Zhao et al. 2017) to all networks. For the ablation study, we train networks for 40K iterations with a batch size of 8 on Cityscapes train set. The results are obtained by the whole test strategy on the validation set. For comparison with SOTA, we train networks with iterations of batch size of 160K and 8 on Cityscapes, 160K and 16 on ADE20K, 30K and 16 on Pascal Context, respectively.