[[arXiv 2305.07895]](https://arxiv.org/pdf/2305.07895.pdf) On the Hidden Mystery of OCR in Large Multimodal Models.
We conducted a comprehensive study of existing publicly available multimodal models, evaluating their performance in text recognition (document text, artistic text, handwritten text, scene text), text-based visual question answering (document text, scene text, and bilingual text), key information extraction (receipts, documents, and nutrition facts) and handwritten mathematical expression recognition. The baseline results showcased in this study could provide a foundational framework for the conception and assessment of innovative strategies targeted at enhancing zero-shot multimodal techniques. Online evaluation DEMO is available at this [link](http://124.220.17.244:7860/).
Results are available in answer_save folder. It should be noted that for BLIP2OPT, when using the inference code on Hugging Face, the accuracy of text recognition is high, but the model outputs nothing for the VQA tasks. Conversely, when using the LAVIS library for inference, the accuracy of text recognition is low, while the VQA accuracy is normal. We believe that the inference process of BLIP2OPT still needs to be optimized. In our experiments, we take the maximum value of the two methods as the final result.
If you want to add a new model, please write its inference function under the folder "models", and update the get_model function in eval.py. An example inference code is as follows：
```Shell
import torch
from PIL import Image
from lavis.models import load_model_and_preprocess
