SphereVAD: A New Paradigm for Training-Free Video Anomaly Detection—Geodesic Reasoning on Hyperspheres

This article introduces SphereVAD, a fully training-free, zero-shot video anomaly detection framework. It leverages intermediate-layer features from pre-trained multimodal large language models (MLLMs), implements anomaly discrimination via von Mises-Fisher (vMF) distribution and geodesic reasoning on hyperspheres, and achieves state-of-the-art performance among training-free methods on three benchmark datasets.

Video Anomaly Detection (VAD) is a key research direction in computer vision, aiming to automatically identify events that deviate from normal patterns in unedited surveillance videos. Traditional methods face a dilemma: either they require large-scale manually labeled data for supervised training, or they need complex task customization for specific scenes. This reliance on training data severely limits their ability to deploy quickly in new scenarios. In practical applications, surveillance systems often need to be put into use immediately in completely new environments without time or resources to collect data or train models. Therefore, developing plug-and-play, training-free anomaly detection methods is an important pursuit in the field.

Researchers found that intermediate-layer features of pre-trained multimodal large language models (MLLMs) already encode rich anomalous semantic information, and naturally have the ability to distinguish between normal and abnormal even without optimization for VAD. Existing methods rely on language output paths, which wastes geometric discriminability. The core innovation of SphereVAD lies in directly utilizing the geometric structure of intermediate-layer features.

The technical solution includes three key components:

1. Frechet Mean Centering: Unfolds the feature distribution, eliminates domain bias, and ensures that features from different videos can be compared in a unified coordinate system;
2. Holistic Scene Attention (HSA): Aggregates multi-video information to enhance feature consistency and help understand normal behavior patterns;
3. vMF-guided Spherical Geodesic Pulling (SGP): Aligns ambiguous segments with directional prototypes on the spherical manifold, and calculates geodesic distance to quantify the degree of deviation.

SphereVAD was evaluated on three video anomaly detection benchmark datasets, with the following results:

• Achieved state-of-the-art performance among all training-free methods;
• Still competitive compared to fully supervised baseline methods;
• Only requires a very small amount of synthetic image calibration.

These results indicate that geometric reasoning can avoid the reliance of traditional methods on large-scale training data while maintaining excellent detection performance.

The zero-shot feature of SphereVAD has extremely high practical value, suitable for scenarios requiring rapid deployment such as temporary event security, emergency response, and resource-constrained edge devices, without the need for model fine-tuning or data collection. In addition, this study reveals that the internal representations of pre-trained large models may contain richer information than explicit outputs, which is expected to inspire researchers in other fields to rethink how to effectively utilize the potential of large models.

SphereVAD transforms anomaly detection into a hypersphere geometric reasoning problem, realizing a fully training-free solution. It not only outperforms similar methods in performance but also demonstrates a new paradigm of 'zero-shot + geometric reasoning'. With the continuous development of multimodal large language models, this idea is expected to show its power in more visual understanding tasks.