ImageNet-9

深度计算机视觉模型依赖于前景对象和图像背景。即使存在正确的前景对象，当图像背景发生变化时，此类模型也经常会做出不正确的预测，并且它们特别容易受到不利选择的背景的影响。例如，官方的预先训练的PyTorch ResNet-50在ImageNet-9上针对对抗性背景进行评估时具有22% 的准确性 (作为参考，始终预测 “狗” 的模型具有11% 的准确性)。Deep computer vision models rely on both foreground objects and image backgrounds. Even when the correct foreground objects are present, such models often produce incorrect predictions when the image background changes, and they are particularly vulnerable to adversarially selected backgrounds. For example, the officially pre-trained PyTorch ResNet-50 achieves only 22% accuracy when evaluated on ImageNet-9 under adversarial backgrounds. For reference, a model that consistently predicts "dog" attains just 11% accuracy.

搜集汇总

背景与挑战

背景概述

ImageNet-9数据集用于评估深度计算机视觉模型对图像背景变化的敏感性，例如在对抗性背景下，预训练模型准确率可能显著下降至22%。该数据集由麻省理工大学于2020年发布，相关资源可在其GitHub页面和论文中获取。