MG
M. Girish Nair
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
Style-based anomaly detection occurs when an action is performed in an abnormal manner. In this paper, we investigate how anomaly detection models generalize when actors, actions or anomaly styles are unseen during training, a setting that is not directly covered by existing benchmarks. To address this, we propose a shared anomaly direction model based on MotionCLIP. The model learns a direction in the motion embedding space that separates normal and anomalous executions and can transfer across different anomaly styles. The model is evaluated on the PersonaMotion dataset using splits with unseen actors, actions and styles. It achieves the highest mean AUROC across the evaluated settings, with particularly strong performance on unseen actors and actions. However, the performance decreases when more anomaly styles are withheld during training. This suggests that a transferable shared anomaly component can be learned, although style-specific variation still remains.
...
Style-based anomaly detection occurs when an action is performed in an abnormal manner. In this paper, we investigate how anomaly detection models generalize when actors, actions or anomaly styles are unseen during training, a setting that is not directly covered by existing benchmarks. To address this, we propose a shared anomaly direction model based on MotionCLIP. The model learns a direction in the motion embedding space that separates normal and anomalous executions and can transfer across different anomaly styles. The model is evaluated on the PersonaMotion dataset using splits with unseen actors, actions and styles. It achieves the highest mean AUROC across the evaluated settings, with particularly strong performance on unseen actors and actions. However, the performance decreases when more anomaly styles are withheld during training. This suggests that a transferable shared anomaly component can be learned, although style-specific variation still remains.