CoSW: Conditional Sample Weighting for Smoke Segmentation with Label Noise

Authors: Lujian Yao, Haitao Zhao, Zhongze Wang, Kaijie Zhao, Jingchao Peng

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

Reproducibility Variable Result LLM Response
Research Type Experimental The experiments show that our approach achieves SOTA performance on both real-world and synthetic noisy smoke segmentation datasets.
Researcher Affiliation Academia Lujian Yao Haitao Zhao Zhongze Wang Kaijie Zhao Jingchao Peng School of Information Science and Engineering East China University of Science and Technology Shanghai, China {lujianyao,zzwang,kjzhao,pjc}@mail.ecust.edu.cn, haitaozhao@ecust.edu.cn
Pseudocode No The paper provides mathematical derivations and outlines a process flow in Figure 2, but it does not include a dedicated pseudocode block or algorithm listing.
Open Source Code No The paper has not yet been open-sourced for data and code.
Open Datasets Yes Due to the visual diversity and blurry edges of smoke, label noise in the large-scale real-world smoke datasets (Smoke Seg [53] and SMOKE5K [51]) is ubiquitous. Hence, we conduct experiments on both datasets as real-world noise evaluation.
Dataset Splits Yes We select 1,000 images from Smoke Seg and carefully re-annotate them to obtain clean labels. Among them, 700 images are used for training, and 300 images are used for validation.
Hardware Specification No The paper does not provide sufficient information on the computer resources, such as the type of comput workers, memory, and time of execution needed to reproduce the experiments.
Software Dependencies No We implement our method on MMSegmentation. Standard color jittering, random cropping, and random flipping are adopted for data augmentation during the training stage.
Experiment Setup Yes We utilize the Adam W optimizer, with learning rate starting at 6e-5 and scheduled according to the polynomial annealing policy. For the Smoke Seg, we crop images to a size of 512 512 for training, while for SMOKE5K, we follow the previous methods and resize the images to 480 480.