跳转至

文章背景与核心概要

在自动驾驶领域，传统的感知系统往往难以应对“词表外”（Out-of-Vocabulary, OOV）物体，例如突然出现的牲畜或道路碎片等未在训练集内出现的物品。传统的闭集检测器往往被迫忽略这些物体，或者以高置信度将其错误分类，这给实际行驶带来了重大的安全隐患。

本文引入了一种开放世界感知层，利用类别无关的区域提议（region proposals）进行分类抽象（taxonomic abstraction）。通过超越扁平化的标签集，该系统能够以分层的方式识别和分类物体。当遇到未知物体时，系统会保守地将其标记为“未知障碍物”（“UNKNOWN OBSTACLE”），而不是强行给出潜在危险的错误分类。研究表明，这种方法消除了高置信度的分类错误，为真实世界的自动驾驶提供了一个稳健的、安全第一的框架。

Open-World Hierarchical Perception: Taxonomic Abstraction over Class-Agnostic Proposals for the Safe Handling of Out-of-Vocabulary Road Objects

Author: Felix Schaller
Date: August 4, 2026
arXiv ID: 2608.07577
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Robotics (cs.RO)

标准自动驾驶感知系统通常难以应对“词表外”（OOV）物体——即未包含在其训练集中的物品，如牲畜或道路碎片。传统的闭集检测器被迫要么忽略这些物体，要么以高置信度对其进行错误分类，从而带来显著的安全风险。

Summary

Standard autonomous driving perception systems often struggle with "out-of-vocabulary" (OOV) objects—items not included in their training sets, such as livestock or road debris. Traditional closed-set detectors are forced to either ignore these objects or misclassify them with high confidence, posing significant safety risks.

本文引入了一种开放世界感知层，该层在类无关区域提议之上利用了分类抽象。通过超越扁平的标签集，系统可以分层识别和分类物体。当物体未知时，系统保守地将其标记为“未知障碍物”，而不是强行进行潜在危险的错误分类。研究表明，这种方法消除了自信的分类错误，为现实世界的自主导航提供了稳健的、安全第一的框架。

This paper introduces an open-world perception layer that utilizes taxonomic abstraction over class-agnostic region proposals. By moving beyond flat label sets, the system can identify and categorize objects hierarchically. When an object is unknown, the system conservatively flags it as an "UNKNOWN OBSTACLE" rather than forcing a potentially dangerous incorrect classification. The study demonstrates that this approach eliminates confident categorical mistakes, providing a robust, safety-first framework for real-world autonomous navigation.

  • 分类抽象： 实现了一个位于类无关提议之上的分层分类层，使系统能够处理传统检测器会漏掉的物体。
  • 安全第一的评估： 进行了基于真实情况的“留类”（leave-classes-out）基准测试。标准的扁平头检测器对 OOV 物体 100% 产生自信的错误，而所提出的分层层发出了零个自信错误的具体标签。
  • 多线索融合： 分析了结合类无关分割、基于外观的分布外（OOD）评分以及单目深度的可行性，得出的结论是没有任何单一的 2D 线索本身是充分的。
  • 保守处理： 系统成功处理了 94% 的 OOV 物体，方法是识别正确的超级类别或将其标记为未知，将安全性置于高特异性（但可能错误）的标签之上。

Key Contributions

  • Taxonomic Abstraction: Implements a hierarchical classification layer that sits atop class-agnostic proposals, allowing the system to handle objects that traditional detectors would miss.
  • Safety-First Evaluation: Conducted a ground-truth "leave-classes-out" benchmark. While a standard flat-head detector produced confident errors 100% of the time on OOV objects, the proposed hierarchical layer emitted zero confident wrong specific labels.
  • Multi-Cue Integration: Analyzed the feasibility of combining class-agnostic segmentation, appearance-based out-of-distribution (OOD) scoring, and monocular depth, concluding that no single 2D cue is sufficient on its own.
  • Conservative Handling: The system successfully handles 94% of OOV objects by either identifying the correct super-category or flagging them as unknown, prioritizing safety over high-specificity (but potentially incorrect) labeling.

Paper Metadata & Resources

license icon

Access