AUCSeg: AUC-oriented Pixel-level Long-tail Semantic Segmentation

Authors: Boyu Han, Qianqian Xu, Zhiyong Yang, Shilong Bao, Peisong Wen, Yangbangyan Jiang, Qingming Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, comprehensive experiments across various benchmarks confirm the effectiveness of our proposed AUCSeg method.
Researcher Affiliation Academia Boyu Han1,2 Qianqian Xu1,3 Zhiyong Yang2 Shilong Bao2 Peisong Wen1,2 Yangbangyan Jiang2 Qingming Huang2,1,4 1 Key Lab. of Intelligent Information Processing, Institute of Computing Technology, CAS 2 School of Computer Science and Tech., University of Chinese Academy of Sciences 3 Peng Cheng Laboratory 4 Key Laboratory of Big Data Mining and Knowledge Management, CAS
Pseudocode Yes Algorithm 1: AUCSeg Algorithm (Short Version)
Open Source Code Yes The code is available at https://github.com/boyuh/AUCSeg.
Open Datasets Yes The experiment includes three benchmark datasets: Cityscapes [19], ADE20K [109], and COCO-Stuff 164K [9].
Dataset Splits Yes The training, validation, and testing set numbers are 2975, 500, and 1525, respectively. (Cityscapes) ...with 20210, 2000, and 3352 images used for training, validation, and testing, respectively. (ADE20K)
Hardware Specification Yes We perform all experiments using mmsegmentation [18] on an NVIDIA 3090 GPU.
Software Dependencies No The paper mentions software like mmsegmentation and AdamW optimizer, but does not provide specific version numbers for these software dependencies (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes We use Adam with Weight Decay (Adam W) [59] optimizer with an initial learning rate of 6e-5 and a weight decay of 0.01. We adopt the poly learning rate policy, where the initial learning rate is multiplied by 1 − iter/max_iter. Moreover, a linear warmup strategy is employed at the beginning of training, allowing the learning rate to increase from 1e-6 to the initial learning rate within 1500 iterations. The batch size is set to 2 for the Cityscapes dataset and 4 for all the other datasets. The total number of iterations is 160000 on Cityscapes and ADE20K and 80000 on COCO-Stuff 164K.