| Property | Value |
|---|---|
| Parameter Count | 28.8M |
| License | MIT |
| Paper | PubTables-1M Paper |
| Downloads | 2.9M+ |
| Tensor Type | F32 |
What is table-transformer-detection?
Table Transformer Detection is a specialized DETR-based model designed for identifying and localizing tables within documents. Developed by Microsoft, this model represents a significant advancement in document understanding, particularly for extracting structured information from unstructured documents.
Implementation Details
The model is built on the DETR (Detection Transformer) architecture with a crucial modification implementing the "normalize before" approach, where layer normalization is applied before self- and cross-attention operations. It's trained on the comprehensive PubTables1M dataset, making it particularly effective for academic and professional document processing.
- Transformer-based architecture using DETR framework
- Modified attention mechanism with pre-normalization
- Optimized for table detection tasks
- Implemented using PyTorch with Safetensors support
Core Capabilities
- Accurate table detection in various document formats
- Processing of complex document layouts
- Integration with inference endpoints for production deployment
- Handling of both simple and complex table structures
Frequently Asked Questions
Q: What makes this model unique?
This model's uniqueness lies in its specialized focus on table detection using the DETR architecture, combined with its training on the extensive PubTables1M dataset. The pre-normalization approach enhances its stability and performance.
Q: What are the recommended use cases?
The model is ideal for automated document processing systems, academic paper analysis, invoice processing, and any application requiring accurate table extraction from documents. It's particularly well-suited for batch processing of large document collections.