跳转至

CogVis:开放词汇表变化检测是否需要为每个查询重新感知场景?

文章背景与核心概要

开放词汇表变化检测(OVCD)通过识别任意语义类别,对于地表监测至关重要。然而,由于时间感知、语义判别和区域验证的相互纠缠,传统方法往往存在结果不稳定和计算冗余的问题。

本文介绍了 CogVis,这是一个新颖的由认知记忆引导的框架,它将 OVCD 重新构想为精简的 感知-记忆-验证(perception-memory-verification) 范式。通过将时间证据与语义类别决策解耦并利用可重用的先验知识,CogVis 在各类基准测试中实现了最先进(SOTA)的性能,同时将推理吞吐量提升了 28.50%。


摘要 (Summary)

Open-Vocabulary Change Detection (OVCD) is crucial for monitoring earth surfaces by recognizing arbitrary semantic categories. However, conventional methods often suffer from unstable results and redundant computations due to the entanglement of temporal perception, semantic discrimination, and region verification.

This paper introduces CogVis, a novel cognitive memory-guided framework that reformulates OVCD into a streamlined perception-memory-verification paradigm. By decoupling temporal evidence from semantic category decisions and leveraging reusable priors, CogVis achieves state-of-the-art performance across diverse benchmarks while improving inference throughput by 28.50%.

开放词汇表变化检测(OVCD)通过识别任意语义类别,对于地表监测至关重要。然而,由于时间感知、语义判别和区域验证的相互纠缠,传统方法往往存在结果不稳定和计算冗余的问题。

本文介绍了 CogVis,这是一个新颖的由认知记忆引导的框架,它将 OVCD 重新构想为精简的 感知-记忆-验证 范式。通过将时间证据与语义类别决策解耦并利用可重用的先验知识,CogVis 在各类基准测试中实现了最先进的性能,同时将推理吞吐量提升了 28.50%。


论文元数据 (Paper Metadata)

  • arXiv ID: arXiv:2608.06150 [cs.AI]
  • Authors: Zijie Wang, Chen Zhong, Wei He
  • Submitted Date: August 6, 2026
  • Primary Subject: Artificial Intelligence (cs.AI)
  • Secondary Subject: Computer Vision and Pattern Recognition (cs.CV)
  • Code Repository: GitHub - KotlinWang/CogVis
  • arXiv ID: arXiv:2608.06150 [cs.AI]
  • 作者: Zijie Wang, Chen Zhong, Wei He
  • 提交日期: 2026年8月6日
  • 主学科: 人工智能 (cs.AI)
  • 次学科: 计算机视觉与模式识别 (cs.CV)
  • 代码仓库: GitHub - KotlinWang/CogVis

核心架构与方法论 (Core Architecture & Methodology)

Inspired by human visual change perception, CogVis breaks down the OVCD pipeline into three core components:

  1. Scene Change Perceptron (SCP): Extracts a reusable, category-agnostic change prior directly from frozen bi-temporal features. This effectively separates category-agnostic temporal perception from high-level semantic decisions, preventing redundant computations across multiple queries.
  2. Semantic Memory Calibrator (SMC): Dynamically estimates an image-query-specific decision threshold to compensate for category-dependent score shifts.
  3. Adaptive Region Filter (ARF): Filters connected candidates utilizing learned semantic, temporal, and structural reliability metrics to yield robust final detections.

受人类视觉变化感知的启发,CogVis 将 OVCD 流程拆解为三个核心组件:

  1. 场景变化感知器(SCP): 直接从冻结的双时相特征中提取可重用的、与类别无关的变化先验。这有效地将与类别无关的时间感知与高层语义决策分离,防止了多个查询之间的冗余计算。
  2. 语义记忆校准器(SMC): 动态估计特定于图像和查询的决策阈值,以补偿类别相关的得分偏移。
  3. 自适应区域滤波器(ARF): 利用学到的语义、时间和结构可靠性指标来过滤连通候选区域,从而产生稳健的最终检测结果。

实验结果 (Experimental Results)

CogVis has been evaluated across seven benchmarks, covering a wide array of tasks: * Semantic Change Detection * Binary Change Localization * Building-Damage Assessment

Key Highlights: * State-of-the-Art Performance: Outperforms existing methods consistently across all evaluated datasets. * Enhanced Efficiency: Avoids repeating category-agnostic temporal perception for every new query, resulting in a 28.50% improvement in inference throughput.

CogVis 已在 七个基准测试 上进行了评估,涵盖了广泛的任务: * 语义变化检测 * 二元变化定位 * 建筑破坏评估

核心亮点: * 最先进的性能: 在所有评估的数据集上均持续优于现有方法。 * 效率提升: 避免了针对每个新查询重复进行与类别无关的时间感知,从而使 推理吞吐量提升了 28.50%