Strictly Causal Streaming Video Anomaly Detection with a Theoretically-Grounded State-Space Core

Forum topic · 小凯 · 2026-08-27

Summary

A recent arXiv paper (2608.24810) by Yogesh Kumar introduces a strictly causal streaming video anomaly detector built on a Mamba-style state space model. Unlike prior approaches that internally buffer clips or windows, this detector maintains a fixed-size state updated in O(1) time and memory per incoming frame, with no lookahead and no clip buffering, making it suitable for edge deployment. The temporal core is a diagonal linear state space recurrence with an input- and state-dependent decay gate, trained self-supervised through causal next-embedding prediction on a frozen visual backbone. A key theoretical contribution is a closed-form relationship linking the recurrence's decay spectrum to detection latency and the shortest anomaly duration that can be reliably captured—addressing the lack of theory connecting temporal memory to latency in prior work. The paper also criticizes existing efficiency benchmarks that rely solely on GPU throughput rather than the edge hardware these methods target. This work is relevant for real-time video surveillance and anomaly detection on resource-constrained devices.

Paper Overview

Field: AI
Author: Yogesh Kumar
Published: 2026-08-25
arXiv: 2608.24810

Abstract

Recent work has applied Mamba-style state space models (SSMs) to video anomaly detection, yet existing approaches still rely on buffering clips or windows internally, lack a theoretical account of how temporal memory relates to detection latency, and benchmark efficiency only through GPU throughput rather than the edge hardware these methods are intended to target.

The paper introduces a strictly causal streaming anomaly detector whose fixed-size state is updated in O(1) time and memory per incoming frame, with no lookahead and no clip buffering. Its temporal core is a diagonal linear state space recurrence with an input- and state-dependent decay gate, trained self-supervised through causal next-embedding prediction on a frozen visual backbone.

Key Contributions

Strict causality: O(1) per-frame time and memory updates with no lookahead or clip buffering, enabling true streaming inference.
State-space temporal core: A diagonal linear recurrence with input- and state-dependent decay gating, in the Mamba style.
Self-supervised training: Causal next-embedding prediction on top of a frozen visual backbone.
Theoretical grounding: A closed-form relationship between the recurrence's decay spectrum, detection latency, and the shortest anomaly duration that can be reliably captured.
Edge-oriented efficiency: Addresses the gap where prior work benchmarks only GPU throughput instead of the edge hardware actually targeted.

---

*Auto-collected on 2026-08-27.*