Multi-pose Anomaly Detection (MAD) 数据集

Multi-pose Anomaly Detection (MAD) 数据集是首个针对多姿态异常检测的评估数据集，由清华大学智能产业研究院开发。该数据集包含20种不同形状和颜色的乐高动物玩具，涵盖超过11,000张高分辨率RGB图像，提供像素级精确的3种异常类型标注。数据集创建过程中，使用了Blender和Ldrew构建3D乐高动物模型，并结合Photoshop手动编辑异常，确保了数据的真实性和多样性。该数据集主要应用于机器视觉领域，旨在解决物体异常检测中由于姿态变化导致的检测难题，推动无姿态限制的异常检测技术发展。

Multi-pose Anomaly Detection (MAD) Dataset is the first evaluation dataset dedicated to multi-pose anomaly detection, developed by the Institute for AI Industry Research of Tsinghua University. This dataset includes 20 LEGO animal toys with varied shapes and colors, comprising over 11,000 high-resolution RGB images, and provides pixel-level precise annotations for three types of anomalies. During the dataset construction, Blender and LDraw were utilized to create 3D models of the LEGO animal toys, and anomalies were manually edited using Photoshop to guarantee the authenticity and diversity of the dataset. This dataset is primarily applied in the field of machine vision, aiming to tackle the detection challenges induced by pose variations in object anomaly detection, and advance the development of pose-unrestricted anomaly detection technologies.

搜集汇总

背景与挑战

背景概述

Multi-pose Anomaly Detection (MAD)数据集是清华大学开发的针对多姿态异常检测的首个评估数据集，包含20种乐高玩具的11,000+高分辨率图像，提供3种像素级异常标注，用于解决物体姿态变化导致的异常检测难题。