Real5-OmniDocBench: Document Images in 5 Real-World Handheld Scenarios

PaddlePaddle released Real5-OmniDocBench in 2026 as a benchmark for document analysis across five real-world scenarios including warping and illumination. The collection contains between 1,000 and 10,000 images, most of which were manually captured using handheld mobile devices to simulate authentic user conditions.

Use Cases

• Evaluating OCR performance on the 'Screen-Photography' subset to handle pixel-grid artifacts
• Testing document dewarping models using the 'Warping' and 'Skew' image categories
• Benchmarking illumination-invariant text recognition on the 'Illumination' subset

Strengths

• Manual acquisition via handheld mobile devices for realistic noise
• Five distinct scenario categories for targeted stress testing
• One-to-one correspondence with OmniDocBench v1.5 for benchmarking

Limitations

• Relatively small scale with 1,000 to 10,000 total images
• Potential geographic bias as the region is listed as US-only
• Lack of detailed column-level metadata in the current distribution

Provenance

Source

PaddlePaddle

Collection Method

Manual acquisition via handheld mobile devices and scanning

Freshness

Last updated March 2026.

Geography

United States