Rethinking Alignment and Uniformity in Unsupervised Image Semantic Segmentation

Authors: Daoan Zhang, Chenming Li, Haoquan Li, Wenjian Huang, Lingyun Huang, Jianguo Zhang

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on multiple semantic segmentation benchmarks show that our unsupervised segmentation framework specializes in catching semantic representations, which outperforms all the unpretrained and even several pretrained methods.Ablation Study Pixel-wise Encoder and Semantic-wise Generator. We first tested the effects of different backbones on the pixel-wise encoder and semantic-wise generator. The results are presented in Table 4.
Researcher Affiliation Collaboration Daoan Zhang1,3, Chenming Li1, Haoquan Li1, Wenjian Huang1, Lingyun Huang3, Jianguo Zhang1, 2* 1Southern University of Science and Technology 2Peng Cheng Laboratory 3Ping An Technology (Shenzhen) Co., Ltd.
Pseudocode No No pseudocode or clearly labeled algorithm block is present in the paper.
Open Source Code No No statement explicitly providing access to source code for the described methodology or a link to a code repository is found in the paper.
Open Datasets Yes The experiments are conducted on five semantic segmentation benchmarks of different scenarios, including COCOStuff-3(Caesar, Uijlings, and Ferrari 2018), COCO-Stuff-15, COCO-Stuff-27, Cityscapes(Cordts et al. 2016) and POTSDAM(Rottensteiner et al. 2014) with various segmentation targets.
Dataset Splits Yes We also apply the widely used Cityscapes dataset, which contains 5,000 images focusing on street scenes, divided into 2,975 and 500 images used for training and validation.
Hardware Specification Yes All datasets are trained for 30 epochs on 4 Tesla V100 32G GPUs with Adam optimizer.
Software Dependencies No All datasets are trained for 30 epochs on 4 Tesla V100 32G GPUs with Adam optimizer. The paper mentions the optimizer but does not specify software dependencies like programming languages or deep learning frameworks with version numbers.
Experiment Setup Yes All datasets are trained for 30 epochs on 4 Tesla V100 32G GPUs with Adam optimizer. The learning rate is set to 1e 4.