TAO-OW

跟踪和检测任何对象，包括在模型训练期间从未见过的对象，是自治系统的一项至关重要但难以捉摸的能力。在现实世界中操作时，对从未见过的物体视而不见的自治代理会带来安全隐患-但这几乎是所有当前系统的工作方式。推进跟踪任何对象的主要障碍之一是，这项任务不难评估。使我们能够对现有工作进行苹果对苹果比较的基准是推进这一重要研究领域的关键第一步。本文解决了这种评估缺陷，并提出了在开放世界环境中检测和跟踪已知和未知物体的景观和评估方法。我们提出了一个新的基准TAO-OW: 跟踪开放世界中的任何对象，分析多对象跟踪中的现有工作，并为此任务构建基线，同时突出未来的挑战。我们希望在多目标跟踪研究中开辟新的前沿，有望使我们更接近可以在现实世界中安全运行的智能系统。

Tracking and detecting any object—including those never encountered during model training—represents a critical yet elusive capability for autonomous systems. When operating in real-world environments, autonomous agents that remain oblivious to unseen objects pose significant safety risks, yet this is the standard mode of operation for nearly all current systems. One of the primary obstacles to advancing any-object tracking is that this task has long lacked robust, standardized evaluation frameworks. Benchmarks that enable apple-to-apple comparisons of existing work represent a crucial first step in advancing this important research area. This paper addresses this critical evaluation gap, and proposes a comprehensive landscape and evaluation methodology for detecting and tracking both known and unknown objects in open-world environments. We introduce a new benchmark, TAO-OW: Tracking Any Object in the Open World, analyze existing research in multi-object tracking, build solid baselines for this task, and highlight key future challenges. We aim to open new frontiers in multi-object tracking research, bringing us closer to intelligent systems that can operate safely in real-world scenarios.

搜集汇总

背景与挑战

背景概述

TAO-OW是一个专注于开放世界中检测和跟踪已知和未知物体的数据集，旨在通过新的基准和基线推动多对象跟踪研究。该数据集由多家知名大学联合发布，适用于自动驾驶等需要安全识别未见物体的智能系统研究。