SmooSeg: Smoothness Prior for Unsupervised Semantic Segmentation

Authors: Mengcheng Lan, Xinjiang Wang, Yiping Ke, Jiaxing Xu, Litong Feng, Wayne Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our Smoo Seg significantly outperforms STEGO in terms of pixel accuracy on three datasets: COCOStuff (+14.9%), Cityscapes (+13.0%), and Potsdam-3 (+5.7%).
Researcher Affiliation Collaboration 1 S-Lab, Nanyang Technological University 2 SCSE, Nanyang Technological University 3 Sense Time Research
Pseudocode Yes Algorithm 1 Smoo Seg: Py Torch-like Pseudocode
Open Source Code Yes https://github.com/mc-lan/Smoo Seg
Open Datasets Yes We test on three datasets. COCOStuff [35] is a scene-centric dataset... Cityscapes [36] is a collection of street scene images... Potsdam-3 [3] is a remote sensing dataset...
Dataset Splits No The paper states training and testing image counts for Potsdam-3, but does not explicitly provide details about a separate validation set split for any of the datasets.
Hardware Specification Yes Our experiments were conducted using Py Torch [37] on an RTX 3090 GPU.
Software Dependencies No The paper mentions 'Py Torch [37]' but does not provide a specific version number for PyTorch or any other software dependencies.
Experiment Setup Yes The exponential moving average (EMA) hyper-parameter is set to α = 0.998. The dimension of the embedding space is D = 64. The temperature is set to τ = 0.1. We use the Adam optimizer [38] with a learning rate of 1 10 4 and 5 10 4 for the projector and predictor, respectively. We set a batch size of 32 for all datasets. We train our model with a total of 3000 iterations for Cityscapes and Potsdam-3 datasets, and 8000 iterations for the COCOStuff dataset.