ObjectNet

“ObjectNet 是直接使用众包收集的图像测试集。一种新的视觉数据集，借鉴了其他科学领域的控制思想。没有训练集，只有一个测试集！让你的视觉系统通过它速度。收集以有意在新背景上显示来自新视点的对象。50,000 个图像测试集，与 ImageNet 相同，具有旋转、背景和视点控制。313 个对象类与 113 个重叠的 ImageNet 性能下降大，您可以从视觉中得到什么现实世界中的系统！对微调和非常困难的迁移学习问题具有鲁棒性”

ObjectNet is an image test set collected directly via crowdsourcing. As a novel visual dataset, it draws on controlled experimental principles from other scientific fields. There is no training split, only a test split! It can be used to evaluate the speed and performance of vision systems. The dataset was collected to intentionally present objects from novel viewpoints against new backgrounds. Comprising 50,000 test images, it matches the scale of ImageNet, with controlled variations in rotation, background, and viewpoint. It covers 313 object categories, 113 of which overlap with those in ImageNet. This dataset can accurately reflect the performance degradation of real-world vision systems, enabling assessment of their actual performance in real-world scenarios, and is robust to fine-tuning tasks and highly challenging transfer learning problems.

搜集汇总

背景与挑战

背景概述

ObjectNet是一个包含50,000张图像的无监督测试集，专注于通过控制旋转、背景和视点来评估视觉系统的性能。该数据集包含313个对象类，与ImageNet有113个重叠类，旨在推动物体识别模型的极限。