Research map / Vision-language and multimodal / Cross-modal learning

Rethinking Reverse Distillation for Multi-Modal Anomaly Detection

AAAI 2024 · Vision-language and multimodal › Cross-modal learning · approach: Multi-modality image fusion · 34 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Multi-Modal Reverse Distillation (MMRD), a frozen siamese multi-modal teacher with parameter-free modality modulation and a student decoder using learned prototypes, for RGB+depth/normal unsupervised anomaly detection and localization.

Problem

Unsupervised industrial anomaly detection based on RGB images misses anomalies that are invisible in color, so extra modalities (depth, surface normals, point clouds) are needed, but knowledge distillation has not been explored for multi-modal anomaly detection.

Approach

Extends reverse distillation to multi-modal inputs: a frozen siamese teacher with shared convolutions and individual BNs fuses RGB and auxiliary-modality features parameter-free (gated by a normalized sigmoid weight) to form distillation targets, while a student decoder generates modality-related priors via learned prototypes and multi-modal interaction to restore those targets.

Key results

• On MVTec 3D-AD with RGB+3D, MMRD reaches mean image AUROC 95.0 and PRO 97.6, above M3DM (94.5/96.4) and AST (93.7/94.6).
• On Eyecandies, MMRD achieves mean image AUROC 89.5 and PRO 90.6 (RGB+3D setting in Table 1b), with the ablation's full model reaching 93.6 PRO on Eyecandies with normals.
• MMRD is about 5.8 GPU hours with 10.2 FPS versus M3DM's 12.6 GPU hours at 0.10 FPS (Table 3).

Measured effects reported in the paper

Surprising findings

• Element-wise addition of teacher features (no adaptive weight, alpha=1) outperforms parameter-free self-attention fusion (CSA/SSA) and a learnable SE module for fusing modalities.
• Using the sum of RGB and auxiliary features to compute the fusion weight works better than using the auxiliary features alone.

Limitations

• Relies on auxiliary modality (depth/normals/point clouds) being available alongside RGB.
• Localization PRO on some Eyecandies configurations and AUPR-based comparisons are not uniformly best; gains on some categories are small.
• Only evaluated on two multi-modal benchmarks; code promised only upon acceptance.

Experimental setup

Datasets: MVTec 3D-AD Eyecandies

Benchmarks: MVTec 3D-AD Eyecandies

Metrics: AUROC AUPR PRO GPU hours FPS

Baselines: RD FPFH PatchCore AST M3DM Eyecandy FD MMFD

Builds on: Reverse Distillation Student-Teacher feature pyramid PatchCore Squeeze-and-Excitation

Closest papers by meaning

• Beyond Single-Modal Boundary: Cross-Modal Anomaly Detection through Visual Prototype and Harmonization (CVPR 2025 · similarity 0.88)
• Unlocking the Potential of Reverse Distillation for Anomaly Detection (AAAI 2025 · similarity 0.87)
• Just Dance with pi! A Poly-modal Inductor for Weakly-supervised Video Anomaly Detection (CVPR 2025 · similarity 0.86)
• Multimodal Industrial Anomaly Detection by Crossmodal Feature Mapping (CVPR 2024 · similarity 0.86)
• Align3D-AD: Cross-Modal Feature Alignment and Dual-Prompt Learning for Zero-shot 3D Anomaly Detection (NeurIPS 2026 · similarity 0.86)
• Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection (CVPR 2025 · similarity 0.86)