đ TF-ID: Table/Figure IDentifier for academic papers
TF-ID is a family of object detection models designed to extract tables and figures from academic papers. It offers multiple versions to meet different needs, providing accurate and efficient extraction capabilities.
đ Quick Start
Use the following code to start using the TF-ID model:
import requests
from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("yifeihu/TF-ID-base", trust_remote_code=True)
processor = AutoProcessor.from_pretrained("yifeihu/TF-ID-base", trust_remote_code=True)
prompt = "<OD>"
url = "https://huggingface.co/yifeihu/TF-ID-base/resolve/main/arxiv_2305_10853_5.png?download=true"
image = Image.open(requests.get(url, stream=True).raw)
inputs = processor(text=prompt, images=image, return_tensors="pt")
generated_ids = model.generate(
input_ids=inputs["input_ids"],
pixel_values=inputs["pixel_values"],
max_new_tokens=1024,
do_sample=False,
num_beams=3
)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
parsed_answer = processor.post_process_generation(generated_text, task="<OD>", image_size=(image.width, image.height))
print(parsed_answer)
To visualize the results, see this tutorial notebook for more details.
⨠Features
- Multiple Model Versions: TF-ID comes in four versions, including models with and without caption text extraction, to meet different application scenarios.
- High Accuracy: Finetuned from microsoft/Florence-2 checkpoints, the models have high accuracy in table and figure extraction.
- Manual Annotation: The training data is manually annotated and checked by humans, ensuring the quality of the data.
đĻ Installation
The code example above shows how to load the model using the transformers
library. You can install the necessary libraries via the following command:
pip install requests pillow transformers
đģ Usage Examples
Basic Usage
import requests
from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("yifeihu/TF-ID-base", trust_remote_code=True)
processor = AutoProcessor.from_pretrained("yifeihu/TF-ID-base", trust_remote_code=True)
prompt = "<OD>"
url = "https://huggingface.co/yifeihu/TF-ID-base/resolve/main/arxiv_2305_10853_5.png?download=true"
image = Image.open(requests.get(url, stream=True).raw)
inputs = processor(text=prompt, images=image, return_tensors="pt")
generated_ids = model.generate(
input_ids=inputs["input_ids"],
pixel_values=inputs["pixel_values"],
max_new_tokens=1024,
do_sample=False,
num_beams=3
)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
parsed_answer = processor.post_process_generation(generated_text, task="<OD>", image_size=(image.width, image.height))
print(parsed_answer)
Advanced Usage
For more advanced usage, such as visualizing the results, refer to this tutorial notebook for detailed instructions.
đ Documentation
Model Summary
TF-ID (Table/Figure IDentifier) is a family of object detection models finetuned to extract tables and figures in academic papers created by Yifei Hu. They come in four versions:
Model |
Model size |
Model Description |
TF-ID-base[HF] |
0.23B |
Extract tables/figures and their caption text |
TF-ID-large[HF] (Recommended) |
0.77B |
Extract tables/figures and their caption text |
TF-ID-base-no-caption[HF] |
0.23B |
Extract tables/figures without caption text |
TF-ID-large-no-caption[HF] (Recommended) |
0.77B |
Extract tables/figures without caption text |
All TF-ID models are finetuned from microsoft/Florence-2 checkpoints. |
|
|
- The models were finetuned with papers from Hugging Face Daily Papers. All bounding boxes are manually annotated and checked by humans.
- TF-ID models take an image of a single paper page as the input, and return bounding boxes for all tables and figures in the given page.
- TF-ID-base and TF-ID-large draw bounding boxes around tables/figures and their caption text.
- TF-ID-base-no-caption and TF-ID-large-no-caption draw bounding boxes around tables/figures without their caption text.
Large models are always recommended!

Object Detection results format:
{'<OD>': {'bboxes': [[x1, y1, x2, y2], ...],
'labels': ['label1', 'label2', ...]} }
Training Code and Dataset
Benchmarks
We tested the models on paper pages outside the training dataset. The papers are a subset of huggingface daily paper.
Correct output - the model draws correct bounding boxes for every table/figure in the given page.
Model |
Total Images |
Correct Output |
Success Rate |
TF-ID-base[HF] |
258 |
251 |
97.29% |
TF-ID-large[HF] |
258 |
253 |
98.06% |
Model |
Total Images |
Correct Output |
Success Rate |
TF-ID-base-no-caption[HF] |
261 |
253 |
96.93% |
TF-ID-large-no-caption[HF] |
261 |
254 |
97.32% |
Depending on the use cases, some "incorrect" output could be totally usable. For example, the model draw two bounding boxes for one figure with two child components.
đ§ Technical Details
- Finetuning: The models are finetuned from microsoft/Florence-2 checkpoints, using papers from Hugging Face Daily Papers as training data.
- Data Annotation: All bounding boxes in the training data are manually annotated and checked by humans, ensuring the accuracy of the data.
- Model Output: The models take an image of a single paper page as the input and return bounding boxes for all tables and figures in the given page.
đ License
This project is licensed under the MIT License.
BibTex and citation info
@misc{TF-ID,
author = {Yifei Hu},
title = {TF-ID: Table/Figure IDentifier for academic papers},
year = {2024},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/ai8hyf/TF-ID}},
}