[arXiv 2305.07895] 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.
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.
Visualization results
| Data file | Size |
|---|---|
| text recognition code:iwyn | 1.37GB |
| STVQA End-to-End Task-3 and Training images | 1.88GB |
| ocrVQA | -- |
| textVQA val set | 6.6GB |
| docVQA Task 1 Validation set | 0.8GB |
| ESTVQA | 5.2GB |
| SROIE | 0.19GB |
| FUNSD | 16MB |
| POIE | 0.43GB |
| HME100K | 0.69GB |
| Google cloud | 9.38GB |
We assume that your symlinked data directory has the following structure:
data
|_ IC13_857
|_ IC15_1811
|_ ...
|_ ESTVQA
|_ textVQA
|_ ...
|_ FUNSD
|_ POIE
eval on all datasets
python eval.py --model_name LLaVA --eval_alleval on one dataset
python eval.py --model_name LLaVA --eval_textVQApython eval.py --model_name LLaVA --eval_ocr --ocr_dataset_name "ct80 IIIT5K"The results will be saved at answer folder.
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：
import torch
from PIL import Image
from lavis.models import load_model_and_preprocess
from ..process import pad_image, resize_image
class lavis:
def __init__(self, model_name, model_type, device) -> None:
model, vis_processors, txt_processors = load_model_and_preprocess(name = model_name, model_type = model_type, is_eval=True, device=device)
self.model_name = model_name
self.model = model
self.vis_processors = vis_processors
self.txt_processors = txt_processors
self.device = device
def generate(self, image, question, name='resize'):
if 'opt' in self.model_name:
prompt = f'Question: {question} Answer:'
elif 't5' in self.model_name:
prompt = f'Question: {question} Short answer:'
else:
prompt = f'Question: {question} Answer:'
image = Image.open(image).convert("RGB")
if name == "pad":
image = pad_image(image, (224,224))
elif name == "resize":
image = resize_image(image, (224,224))
image = self.vis_processors["eval"](image).unsqueeze(0).to(self.device)
prompt = self.txt_processors["eval"](prompt)
answer = self.model.predict_answers(samples={"image": image, "text_input": prompt}, inference_method="generate", max_len=48, min_len=1)[0]
return answer