OpenLORIS-object
收藏资源简介：
计算机视觉最近的突破得益于用于训练的大型代表性数据集（例如 ImageNet 和 COCO）的可用性。然而，机器人视觉对应用从这些标准计算机视觉数据集开发的视觉算法提出了独特的挑战，因为它们对一组固定任务的不变分布的隐含假设。由于计算、存储和有时隐私问题，每次有新任务可用时完全重新训练模型是不可行的，而幼稚的增量策略已被证明会遭受灾难性遗忘的困扰。对于机器人来说，使用自适应视觉感知系统在开放和有害条件下连续运行至关重要，其中终身学习是一项基本能力。然而，很少有数据集和基准可用于评估和比较新兴技术。为了填补这一空白，我们提供了一个新的终身机器人视觉数据集（“OpenLORIS-Object”），通过安装在移动机器人上的 RGB-D 相机收集。该数据集嵌入了机器人在现实生活应用中面临的挑战，并为验证终身对象识别算法提供了新的基准。此外，我们还提供了 9 种最先进的终身学习算法的测试平台。它们中的每一个都涉及 48 个任务，在 OpenLORIS-Object 数据集上具有 4 个评估指标。
Recent breakthroughs in computer vision have benefited from the availability of large-scale representative training datasets such as ImageNet and COCO. However, robotic vision poses unique challenges for visual algorithms developed from these standard computer vision datasets, as these algorithms implicitly assume an invariant distribution over a fixed set of tasks. Given computational, storage, and occasionally privacy constraints, fully retraining models each time a new task is introduced is infeasible, and naive incremental learning strategies have been shown to suffer from catastrophic forgetting. For robotic systems, it is critical to deploy adaptive visual perception systems that operate continuously in open and harsh environments, where lifelong learning is a fundamental capability. However, very few datasets and benchmarks exist for evaluating and comparing emerging lifelong learning technologies. To address this gap, we introduce a new lifelong robotic vision dataset, "OpenLORIS-Object", which was collected using an RGB-D camera mounted on a mobile robot. This dataset encapsulates the real-world challenges encountered by robots and serves as a novel benchmark for validating lifelong object recognition algorithms. Additionally, we provide a testbed containing 9 state-of-the-art lifelong learning algorithms. Each of these algorithms is evaluated across 48 tasks using 4 distinct evaluation metrics on the OpenLORIS-Object dataset.
