Zero-Shot Learning in Industrial Scenarios: New Large-Scale Benchmark, Challenges and Baseline

Researchers introduce MMIO, a large-scale industrial dataset with 80K+ samples, and RTVP, a refined prompt method for zero-shot defect detection in manufacturing. The work addresses the gap between general-purpose Large Visual Language Models and industrial applications, achieving state-of-the-art performance through improved text-visual prompt interactions and domain adaptation.

This research tackles a critical challenge in applying artificial intelligence to industrial manufacturing environments. While Large Visual Language Models have demonstrated impressive capabilities in general computer vision tasks, their performance degrades significantly in industrial settings due to fundamental differences between natural and manufactured scenes—controlled lighting, repetitive patterns, and specific defect types that don't appear in natural image datasets. The introduction of MMIO represents a meaningful contribution to an underserved segment of AI development.

The broader context reflects growing recognition that general-purpose AI models require domain-specific refinement for enterprise deployment. Manufacturing and industrial inspection represent substantial economic value; defect detection alone accounts for significant operational costs across automotive, electronics, and consumer goods sectors. Current approaches either rely on expensive hand-labeled datasets or struggle with the adaptability required for diverse production environments.

The RTVP methodology's innovation centers on automatic visual prompt generation and text-visual interaction modeling, moving beyond simple prompt engineering. This technical advancement has immediate practical implications for manufacturers seeking automated quality control solutions without extensive model retraining. The achieved performance metrics (42.2% AP in zero-shot scenarios) demonstrate competitive viability compared to traditional computer vision approaches.

For the AI and manufacturing technology sectors, this work signals accelerating convergence toward practical industrial AI solutions. Future development will likely focus on expanding MMIO across additional manufacturing domains and optimizing inference speed for real-time production line deployment. The open dataset nature positions it as infrastructure for broader industrial AI ecosystem development.

• →MMIO dataset provides 80K+ industrial samples across 18 subcategories, filling a critical gap in domain-specific AI training data
• →RTVP method achieves state-of-the-art zero-shot defect detection through automatic visual prompt generation and text-visual interaction modeling
• →Expert-guided domain adaptation mechanism enhances large model generalization in controlled industrial environments
• →Open dataset availability accelerates development of practical industrial AI solutions for manufacturing quality control
• →Performance metrics demonstrate feasibility of reducing dependency on extensive labeled datasets for industrial applications