从噪声监督中学习：面向弱监督视频异常检测的去噪-纠偏框架
Learning from Noisy Supervision: A Denoising-Debiasing Framework for Weakly Supervised Video Anomaly Detection
把「高训练损失的样本即噪声」这一观察做成动态丢弃率，再用视觉语言模型对被丢弃片段做二次判别，可作为一种即插即用的去噪-纠偏模块迁移到任何基于 MIL 的视频时序定位任务（如弱监督动作定位、视频级标签的时序 grounding）。
原文依据 We introduce a dynamic drop rate to adaptively filter out suspected noisy samples during training, based on the observation that noisy samples incur higher training losses.
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论弱监督视频异常检测（WS-VAD）旨在仅使用视频级标签定位帧级异常。该任务通常被建模为多实例学习（MIL）范式，将每个视频视为一个由片段构成的包，从而在无需额外信息的情况下取得稳健性能。然而，现有方法往往难以应对带噪的监督信号。异常包中的正常片段常因异常分数不准确而被误判为异常。这些被误分类的实例充当噪声样本，引入虚假监督，阻碍对真实异常模式的学习。在本工作中，我们提出 D^2 MIL，一个位于多实例学习范式内的去噪-纠偏框架，旨在抑制噪声并提升异常判别能力。我们的方法集成了两个关键组件：（1）去噪模块：基于噪声样本会产生更高训练损失的观察，我们引入动态丢弃率，在训练过程中自适应地过滤掉疑似噪声样本。（2）纠偏模块：我们利用视觉语言模型对被丢弃的样本进行重新评估。由于这些样本与噪声样本相似但难以被模型识别，该模块能够找回可能被误删的有价值异常实例。D^2 MIL 是一种通用的去噪策略，可集成到任何基于 MIL 的方法中。我们在三个基准数据集（ShanghaiTech、UCF-Crime 和 MSAD）上的大量实验表明，D^2 MIL 与多种 MIL 框架兼容，并能持续提升其性能。
英文原文摘要 1627 字符
Weakly supervised video anomaly detection (WS-VAD) aims to localize frame-level anomalies using only video-level labels. This task is typically formulated within a multiple instance learning (MIL) paradigm, where each video is treated as a bag of snippets, achieving robust performance without requiring additional information.However, existing methods often struggle with noisy supervision signals. Normal snippets within abnormal bags are frequently misclassified as anomalies due to inaccurate anomaly scores. These misclassified instances act as noisy samples, introducing false supervision that hinders the learning of true anomaly patterns.In this work, we introduce D^ 2 MIL, a Denoising-Debiasing framework within the Multiple Instance Learning paradigm designed to suppress noise and improve anomaly discrimination. Our approach integrates two key components:(1) Denoising Module: We introduce a dynamic drop rate to adaptively filter out suspected noisy samples during training, based on the observation that noisy samples incur higher training losses. (2) Debiasing Module: We leverage a vision-language model to re-evaluate the discarded samples. This recovers potentially valuable abnormal instances that were mistakenly removed, as they are similar to noisy samples but difficult for the model to recognize. D^ 2 MIL is a general purpose denoising strategy that can be integrated into any MIL-based method. Our extensive experiments on the three benchmark datasets (ShanghaiTech, UCF-Crime, and MSAD) demonstrate that D^ 2 MIL is compatible with diverse MIL frameworks and consistently enhances their performance.