MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning

Authors: Bin-Bin Gao

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

Reproducibility Variable Result LLM Response
Research Type Experimental We comprehensively evaluate Meta UAS on three industrial anomaly segmentation benchmarks, MVTec [4], Vis A [78] and Goods [74]. Our method effectively and efficiently segments any anomalies with only one normal image prompt and enjoys trainingfree without guidance from language. Our Meta UAS significantly outperforms previous zero-shot, few-shot, and even full-shot anomaly segmentation methods. Tables 1 and 2 present the comparison results of Meta UAS with the above-mentioned competing methods in generalization and efficiency, respectively.
Researcher Affiliation Industry Bin-Bin Gao Tencent You Tu Lab, Shenzhen, China csgaobb@gmail.com
Pseudocode No The paper describes the Meta UAS framework, including encoder, feature alignment module, and decoder, explaining their functionalities and interactions within the text. However, there is no explicitly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes Code and Models: https://github.com/gaobb/Meta UAS
Open Datasets Yes Following previous works, we comprehensively evaluate Meta UAS on three industrial anomaly segmentation benchmarks, MVTec [4], Vis A [78] and Goods [74]. Specifically, the change segmentation dataset is synthesized using the randomly selected 60,000 images from the MS-COCO training set.
Dataset Splits Yes Then, all these 60,000 samples are divided into training and validation sets with a ratio of 0.95:0.05.
Hardware Specification Yes The evaluation is performed on one V100 GPU with batch size 32.
Software Dependencies No We conduct experiments based on the open-source framework Py Torch.
Experiment Setup Yes The model is trained with 30 epochs on 8 Tesla V100 GPUs with batch size 128. We freeze the encoder and optimize the feature alignment module, the decoder, and the segmentation head with Adam W [25] using weight decay 0.0005 and learning rate 0.0001.