SSA-Seg: Semantic and Spatial Adaptive Pixel-level Classifier for Semantic Segmentation

Authors: Xiaowen Ma, Zhen-Liang Ni, Xinghao Chen

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on three publicly available benchmarks show that the proposed SSA-Seg significantly improves the segmentation performance of the baseline models with only a minimal increase in computational cost.
Researcher Affiliation Collaboration Xiaowen Ma1,2 , Zhenliang Ni1 , Xinghao Chen1 1Huawei Noah s Ark Lab 2Zhejiang University
Pseudocode No The paper describes its methods in text and figures but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes https://github.com/xwmaxwma/SSA-Seg. [...] We provide all the code and configuration files in order to reproduce the experiments in the paper.
Open Datasets Yes We perform experiments on the ADE20K [61], PASCAL-Context [39] and COCO-Stuff-10K [1] datasets. [...] The datasets are publicly available and can be downloaded.
Dataset Splits Yes For ADE20K and COCO-Stuff-10K, we have a cropping size of 512 512, while for PASCAL-Context, we have a cropping size of 480 480. In addition, the batch size of all datasets is 16, and the total iterations for ADE20K, COCO-Stuff-10K and PASCAL-Context number are 160k, 80k and 80k, respectively. [...] The training set, validation set, and the test set contain 20210, 2000, and 3352 images respectively.
Hardware Specification Yes The latency (ms) is calculated on the input size of 512 512 on V100 GPU.
Software Dependencies Yes We use MMSegmentation [12] and follow the common training settings.
Experiment Setup Yes During training, we apply data enhancement sequentially by random horizontal flipping, random resizing with a scale between 0.5 and 2.0, and random cropping. For ADE20K and COCO-Stuff-10K, we have a cropping size of 512 512, while for PASCAL-Context, we have a cropping size of 480 480. In addition, the batch size of all datasets is 16, and the total iterations for ADE20K, COCO-Stuff-10K and PASCAL-Context number are 160k, 80k and 80k, respectively. [...] In the implementation of this paper, λr, λp, λs are all set to 1, and the edge size of boundary is set to 4.