Figure 1. Performance, speed and memory occupancy of Multimodal Anomaly Detection methods.

light & fast & better performance than M3DM, cvpr23

unsupervised multimodal AD: This method, M3DM, BTF.

trained on nominal samples

AST: flow + distillation.

• core idea
• Feature Extraction
• crossmodal feature mapping: $M_{2D->3D}$ & $M_{3D->2D}$
• Training
• Aggregation
• Layers Pruning

1. 应该蒸馏出两个参数量极低的F2D’和F3D’，确保强大的特征提取不会处理好异常？这个由M2D→3D & M3D→2D保证了。要保持F2D, F3D的提取能力，即提出丰富的特征。

Experiments

few shot

5-shot: 5 normal images.