NP-SemiSeg: When Neural Processes meet Semi-Supervised Semantic Segmentation

Authors: Jianfeng Wang, Daniela Massiceti, Xiaolin Hu, Vladimir Pavlovic, Thomas Lukasiewicz

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally evaluated NP-Semi Seg on the public benchmarks PASCAL VOC 2012 and Cityscapes, with different training settings, and the results verify its effectiveness.
Researcher Affiliation Collaboration 1Department of Computer Science, University of Oxford, UK. 2Microsoft Research, Cambridge, UK. 3Department of Computer Science and Technology, Tsinghua University, Beijing, China. 4Department of Computer Science, Rutgers University, New Jersey, USA. 5Vienna University of Technology, Austria.
Pseudocode No The paper does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes The source code is available at: https://github.com/Jianf-Wang/NP-Semi Seg.
Open Datasets Yes We tested our models on two public segmentation benchmarks, namely, Cityscapes (Cordts et al., 2016) and PASCAL VOC 2012 (Everingham et al., 2010)... we used coarsely-labeled 9,118 images from the Segmentation Boundary dataset (SBD) (Hariharan et al., 2011) as additional training data...
Dataset Splits Yes Cityscapes is an urban scene understanding dataset containing 2,975 training images with fine-annotated masks and 500 validation images. PASCAL VOC 2012... There are 1,464 and 1,449 images in the training set and the validation set, respectively.
Hardware Specification Yes All experiments are conducted on Ge Force RTX 3090 GPUs.
Software Dependencies No The paper mentions software components like 'Res Net-50' and 'PyTorch' but does not provide specific version numbers for these or other libraries.
Experiment Setup Yes On the PASCAL VOC 2012 dataset, the training crop size is set to 480 480, and those frameworks with NP-Semi Seg are trained with 0.001 learning rate and 12 batch size. On the Cityscapes dataset, the training crop size is set to 580 580, and we used 0.005 learning rate and 8 batch size for training... The hyper-parameters of NP-Semi Seg include the length of each memory bank (Q), the coefficient λkl, the number of latent maps T. We followed NP-Match to set Q = 2560 for all memory banks. T was set to 5 at both the training phase and the testing phase. The coefficient λkl is set to 0.005.