Dynamic Sampling Network for Semantic Segmentation
Authors: Bin Fu, Junjun He, Zhengfu Zhang, Yu Qiao10794-10801
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We incorporate our proposed CGDS module into Dynamic Sampling Network (DSNet) and perform extensive experiments on segmentation datasets. Experimental results show that our CGDS significantly improves semantic segmentation performance and achieves state-of-the-art performance on PASCAL VOC 2012 and ADE20K datasets. |
| Researcher Affiliation | Collaboration | 1Shen Zhen Key Lab of Computer Vision and Pattern Recognition, SIAT-Sense Time Joint Lab, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences 2SIAT Branch, Shenzhen Institute of Artificial Intelligence and Robotics for Society 3University of Chinese Academy of Science {bin.fu, jj.he, zf.zhang, yu.qiao}@siat.ac.cn |
| Pseudocode | No | The paper includes architectural diagrams (Figure 2, Figure 3) but no explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The codes will become publicly available after publication. |
| Open Datasets | Yes | We incorporate our CGDS module into Dynamic Sampling Network (DSNet) and perform extensive experiments on PASCAL VOC 2012 (Everingham et al. 2010), ADE20K (Zhou et al. 2017) and PASCAL Context (Mottaghi et al. 2014) datasets. |
| Dataset Splits | Yes | The original dataset contains 1, 464 images for training, 1, 449 for validation, and 1, 456 for test. The training set has been augmented to 10, 582 images by extra annotations from (Hariharan et al. 2015) and thus we use this augmented training set in our experiments. It contains 150 semantic classes for scene parsing, with 20, 210 images for training, 2, 000 images for validation and 3, 351 images for testing. |
| Hardware Specification | No | No specific hardware details (e.g., exact GPU/CPU models, memory) were mentioned for running the experiments. |
| Software Dependencies | No | All experiments are performed on the Pytorch (Paszke et al. 2017) platform. No specific version numbers for Pytorch or other libraries are provided. |
| Experiment Setup | Yes | During the training process, we employ the poly learning rate policy lr = initial lr 1 iter total iter 0.9 (Chen et al. 2017; Zhang et al. 2018) for the Res Net backbone of our DSNet where the initial learning rate is 0.01 for PASCAL VOC 2012 and ADE20K datasets, 0.005 for PASCAL Context dataset. While the learning rate for the remaining parts of our DSNet have been setted as 0.1 lr. The network is optimized by stochastic gradient descent (SGD) (Bottou 2010) for 80 epochs on PASCAL VOC 2012 and PASCAL Context datasets, for 150 epochs on ADE20K dataset with momentum 0.9 and weight decay 0.0001. We set the batch size as 32 for PASCAL VOC 2012 and PASCAL Context datasets, 24 for ADE20K dataset. The crop size of input images is chosen as 512 512 for PASCAL VOC 2012 and PASCAL Context datasets, 576 576 for ADE20K dataset. |