DAW: Exploring the Better Weighting Function for Semi-supervised Semantic Segmentation

Authors: Rui Sun, Huayu Mai, Tianzhu Zhang, Feng Wu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results on multiple benchmarks including mitochondria segmentation demonstrate that DAW performs favorably against state-of-the-art methods. and Extensive experiments on mainstream benchmarks demonstrate that our method performs favorably against state-of-the-art semi-supervised semantic segmentation methods, proving that it can better exploit unlabeled data. and sections like 3 Experiments, 3.1 Experimental Setup, 3.2 Comparison with State-of-the-art Methods, 3.3 Ablation Study and Analysis, including multiple tables with quantitative results.
Researcher Affiliation Academia Rui Sun1 Huayu Mai1 Tianzhu Zhang1,2 Feng Wu1,2 1Deep Space Exploration Laboratory/School of Information Science and Technology, University of Science and Technology of China 2Institute of Artificial Intelligence, Hefei Comprehensive National Science Center
Pseudocode Yes The algorithm flow is shown in the supplementary material. (Section 2.4)
Open Source Code Yes Code is available at https://github.com/yuisuen/DAW.
Open Datasets Yes Datasets: (1) PASCAL VOC 2012 [29] is an object-centric semantic segmentation dataset, containing 21 classes with 1,464 and 1,449 finely annotated images for training and validation, respectively. Some researches [30, 19] augment the original training set (e.g., classic) by incorporating the coarsely annotated images in SBD [31], obtaining a training set (e.g., blender) with 10,582 labeled samples. (2) Cityscapes [32]is an urban scene understanding dataset with 2,975 images for training and 500 images for validation.
Dataset Splits Yes PASCAL VOC 2012 [...] 1,464 and 1,449 finely annotated images for training and validation, respectively. and Cityscapes [...] 2,975 images for training and 500 images for validation.
Hardware Specification Yes We train the model for 80 epochs on PASCAL and 240 epochs on Cityscapes, using 8 NVIDIA Ge Force RTX 3090 GPUs.
Software Dependencies No The paper mentions software components like 'SGD as the optimizer', 'ResNet as our backbone', and 'DeepLabv3+ as the decoder', but it does not specify version numbers for any of these or other software dependencies.
Experiment Setup Yes We set the crop size as 513x513 for PASCAL and 801x801 for Cityscapes, respectively. For both datasets, we adopt SGD as the optimizer with the same batch size of 16 and different initial learing rate, which is set as 0.001 and 0.005 for PASCAL and Cityscapes. We use the polynomial policy to dynamically decay the learning rate along the whole training... We train the model for 80 epochs on PASCAL and 240 epochs on Cityscapes. (from Section 3.1 Implementation Details) and Table 5 for 'momentum'.