Towards Generic Anomaly Detection and Understanding: Large-scale...

Anomaly detection is a crucial task across different domains and data types. However, existing anomaly detection models are often designed for specific domains and modalities. This study explores...

https://arxiv.org/abs/2311.02782

📝 主旨内容

自然语言，类别信息，人类先验，zero/one-shot

Regrettably, GPT-4V does not currently have the capability to directly produce prediction masks. Some methods have attempted to leverage GPT-4V by prompting it to provide bounding boxes [101, 97]. However, this approach appears to be imprecise and poses challenges for GPT-4V. In contrast, the approach presented by SoM [100] involves utilizing SAM [50] to generate visual prompts [81, 50], which are presented in numbered markers. This visual prompting technique shifts the localization task from a pixel-level mask prediction task to a mask-level classification task, effectively reducing the associated complexities and increasing localization precision.(遗憾的是，GPT-4V目前还没有能力直接产生预测掩码。一些方法试图利用GPT-4V，促使它提供边界框[101, 97]。然而，这种方法似乎是不精确的，给GPT-4V带来了挑战。相比之下，SoM[100]提出的方法涉及利用SAM[50]来生成视觉提示[81, 50]，这些提示以编号标记的形式呈现。这种视觉提示技术将定位任务从像素级的掩码预测任务转移到掩码级的分类任务，有效地减少了相关的复杂性，提高了定位精度。)

部分实验复现

细节

实现图像异常定位

点云的输入

上下文的选取

医学——伦理限制

🤗 总结归纳

📎 参考文章

通用异常检测新曙光：华科大等揭秘GPT-4V的全方位异常检测表现

异常检测任务旨在识别明显偏离正常数据分布的异常值，在工业检验、医学诊断、视频监控和欺诈检测等多个领域都发挥了重要作用。

https://www.jiqizhixin.com/articles/2023-11-13