NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
Zero-Shot Learning in Industrial Scenarios: New Large-Scale Benchmark, Challenges and Baseline
TL;DR
Existing Large Visual Language Models (LVLMs) struggle with industrial applications due to data scarcity and irrelevant prompts. To address this, a new dataset called the Multi-Modal Industrial Open Dataset (MMIO) and a Refined Text-Visual Prompt (RTVP) method were developed for zero-shot industrial defect detection.
✦ Why It Matters
Engineers can leverage the MMIO dataset and RTVP method to enhance defect detection in industrial applications.
Key Takeaways
- Utilize the MMIO dataset for training models in industrial contexts.
- Implement RTVP to improve object segmentation in industrial images.
- Explore expert-guided domain adaptation for better model generalization.
- Achieve higher precision in defect detection tasks using state-of-the-art methods.
Full Summary
Large Visual Language Models (LVLMs) have shown great success in various vision tasks, but their application in industrial settings is limited due to significant differences between industrial and natural scenes. This research introduces the Multi-Modal Industrial Open Dataset (MMIO), which consists of over 80,000 samples across diverse industrial categories, making it the first large-scale dataset for zero-shot learning in industrial contexts.
Additionally, the study presents a Refined Text-Visual Prompt (RTVP) that enhances the performance of LVLMs by automatically generating visual prompts from images and incorporating text-visual interactions. The RTVP employs an expert-guided domain adaptation mechanism and is based on the Mobile-SAM model, improving the generalization of large models in industrial scenarios.
Results indicate that RTVP achieved state-of-the-art performance with 42.2% average precision (AP) in zero-shot tasks and 24.7% AP in closed scenes. This work provides valuable resources and methodologies for future research and applications in industrial defect detection.