Transmission-Guided Bayesian Generative Model for Smoke Segmentation

Authors: Siyuan Yan, Jing Zhang, Nick Barnes3009-3017

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on benchmark testing datasets illustrate that our model achieves both accurate predictions and reliable uncertainty maps representing model ignorance about its prediction.
Researcher Affiliation Academia The Australian National University {siyuan.yan, jing.zhang, nick.barnes}@anu.edu.au
Pseudocode No No pseudocode or algorithm blocks were found in the paper. The method is described in text and through network architecture diagrams.
Open Source Code Yes Our code and dataset are publicly available at: https://github.com/redlessme/Transmission-BVM.
Open Datasets Yes Our code and dataset are publicly available at: https://github.com/redlessme/Transmission-BVM. ... To promote the development of this field, we provide the first high-quality, large-scale smoke dataset with 5,000 training images and 400 testing images, where the training set consists of 1,400 real life images and 4,000 synthetic images with pixel-wise annotation, and the test images are all real life images.
Dataset Splits No The paper specifies training and testing sets (5000 training images, 400 testing images) and discusses evaluation on these. However, it does not explicitly define a separate validation set split or how validation was performed to tune hyperparameters during training. Phrases like 'training dataset' and 'testing images' are used, but a distinct 'validation' split is not specified.
Hardware Specification Yes It took about three days of training on SYN70K and 15 hours on SMOKE5K with batch size 6 using a single NVIDIA Ge Force RTX 2080Ti GPU.
Software Dependencies No The paper mentions using "Py Torch" but does not specify a version number or other key software components with their versions (e.g., Python version, CUDA version, specific libraries used with versions).
Experiment Setup Yes Each image is re-scaled to 480 480. Empirically, we set the dimension of the latent space z as 8. The learning rates of the generator Pθ(y|x, z) and the uncertainty estimation network gγ are initialized to 2.5e-5 and 1.5e-5, respectively. We use the Adam optimizer and decrease the learning rate 0.8 after 40 epochs with a maximum epoch of 50. We adopt Res Net50 backbone to the encoder of Pθ(y|x, z), which is initialized with parameters trained for image classification, and the other newly added layers are initialized with the Py Torch default initialization strategy.