nvidia/nemotron-table-structure-v1

Nemotron Table Structure v1

Model Overview

[image]

Preview of the model output on the example image.

The input of this model is expected to be a table image. You can use the Nemotron Page Element v3 to detect and crop such images.

Description

The Nemotron Table Structure v1 model is a specialized object detection model designed to identify and extract the structure of tables in images. Based on YOLOX, an anchor-free version of YOLO (You Only Look Once), this model combines a simpler architecture with enhanced performance. While the underlying technology builds upon work from Megvii Technology, we developed our own base model through complete retraining rather than using pre-trained weights.

The model excels at detecting and localizing the fundamental structural elements within tables. Through careful fine-tuning, it can accurately identify and delineate three key components within tables:

1. 1.Individual cells (including merged cells)
2. 2.Rows
3. 3.Columns

This specialized focus on table structure enables precise decomposition of complex tables into their constituent parts, forming the foundation for downstream retrieval tasks. This model helps convert tables into the markdown format which can improve retrieval accuracy.

This model is ready for commercial/non-commercial use.

We are excited to announce the open sourcing of this commercial model. For users interested in deploying this model in production environments, it is also available via the model API in NVIDIA Inference Microservices (NIM) at nemotron-table-structure-v1.

License/Terms of use

The use of this model is governed by the NVIDIA Open Model License Agreement and the use of the post-processing scripts are licensed under Apache 2.0.

Team

• —Theo Viel
• —Bo Liu
• —Darragh Hanley
• —Even Oldridge

Correspondence to Theo Viel (tviel@nvidia.com) and Bo Liu (boli@nvidia.com)

Deployment Geography

Global

Use Case

The Nemotron Table Structure v1 model specializes in analyzing images containing tables by:

• —Detecting and extracting table structure elements (rows, columns, and cells)
• —Providing precise location information for each detected element
• —Supporting downstream tasks like table analysis and data extraction

The model is designed to work in conjunction with OCR (Optical Character Recognition) systems to:

1. 1.Identify the structural layout of tables
2. 2.Preserve the relationships between table elements
3. 3.Enable accurate extraction of tabular data from images

Ideal for:

• —Document processing systems
• —Automated data extraction pipelines
• —Digital content management solutions
• —Business intelligence applications

Release Date

10/23/2025 via https://huggingface.co/nvidia/nemotron-table-structure-v1

References

• —YOLOX paper: https://arxiv.org/abs/2107.08430
• —YOLOX repo: https://github.com/Megvii-BaseDetection/YOLOX
• —Technical blog: https://developer.nvidia.com/blog/approaches-to-pdf-data-extraction-for-information-retrieval/

Model Architecture

Architecture Type: YOLOX <br> Network Architecture: DarkNet53 Backbone \+ FPN Decoupled head (one 1x1 convolution \+ 2 parallel 3x3 convolutions (one for the classification and one for the bounding box prediction). YOLOX is a single-stage object detector that improves on Yolo-v3. <br> This model was developed based on the Yolo architecture <br> Number of model parameters: 5.4e7 <br>

Input

Input Type(s): Image <br> Input Format(s): Red, Green, Blue (RGB) <br> Input Parameters: Two-Dimensional (2D)<br> Other Properties Related to Input: Image size resized to (1024, 1024)

Output

Output Type(s): Array <br> Output Format: A dictionary of dictionaries containing np.ndarray objects. The outer dictionary has entries for each sample (page), and the inner dictionary contains a list of dictionaries, each with a bounding box (np.ndarray), class label, and confidence score for that page. <br> Output Parameters: One-Dimensional (1D) <br> Other Properties Related to Output: The output contains bounding boxes, detection confidence scores, and object classes (cell, row, column). The thresholds used for non-maximum suppression are conf_thresh = 0.01 and iou_thresh = 0.25 Output Classes: <br>

• —Cell
• —Table cell
• —Row
• —Table row
• —Column
• —Table column

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Usage

The model requires torch, and the custom code available in this repository.

1. 1.Clone the repository

• —Make sure git-lfs is installed (https://git-lfs.com)

git lfs install

• —Using https

git clone https://huggingface.co/nvidia/nemotron-table-structure-v1

• —Or using ssh

git clone git@hf.co:nvidia/nemotron-table-structure-v1

Optional: This can be installed as a package using pip

cd nemotron-table-structure-v1
pip install -e .

1. 1.Run the model using the following code:

import torch
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image

from nemotron_table_structure_v1 import (
    define_model,
    plot_sample,
    postprocess_preds_table_structure,
    reformat_for_plotting,
)

# Load image
path = "./example.png"
img = Image.open(path).convert("RGB")
img = np.array(img)

# Load model
model = define_model("table_structure_v1")

# Inference
with torch.inference_mode():
    x = model.preprocess(img)
    preds = model(x, img.shape)[0]

# Post-processing
boxes, labels, scores = postprocess_preds_table_structure(preds, model.threshold, model.labels)

# Plot
boxes_plot, confs = reformat_for_plotting(boxes, labels, scores, img.shape, model.num_classes)

plt.figure(figsize=(30, 15))
for i in range(1, 4):
    boxes_plot_c = [b if j == i else [] for j, b in enumerate(boxes_plot)]
    confs_c = [c if j == i else [] for j, c in enumerate(confs)]

    plt.subplot(1, 3, i)
    plt.title(model.labels[i])
    plot_sample(img, boxes_plot_c, confs_c, labels=model.labels, show_text=False)
plt.show()

Note that this repository only provides minimal code to infer the model. If you wish to do additional training, refer to the original repo.

1. 1.Advanced post-processing

Additional post-processing might be required to use the model as part of a data extraction pipeline. We show how to use the model as part of a table to text pipeline alongside with the Nemotron OCR in the notebook Demo.ipynb.

Disclaimer: We are aware of some issues with the model, and will provide a v2 with improved performance in the future which addresses the following issues:

• —The model appears to be less confident in detecting cells in the bottom of the table, which sometimes results in missed cells.
• —Add an extra class for table titles
• —Add support for non full-page tables

<!---

Software Integration

Runtime Engine(s):

• —Nemotron Page Elements v3 NIM

Supported Hardware Microarchitecture Compatibility [List in Alphabetic Order]:

• —NVIDIA Ampere
• —NVIDIA Hopper
• —NVIDIA Lovelace

Preferred/Supported Operating System(s):

• —Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment. This AI model can be embedded as an Application Programming Interface (API) call into the software environment described above. --->

Model Version(s):

• —nemotron-table-structure-v1

Training and Evaluation Datasets:

Training Dataset

Data Modality: Image <br> Image Training Data Size: Less than a Million Images <br> Data collection method by dataset: Automated <br> Labeling method by dataset: Automated <br> Pretraining (by NVIDIA): 118,287 images of the COCO train2017 dataset <br> Finetuning (by NVIDIA): 23,977 images from Digital Corpora dataset, with annotations from Azure AI Document Intelligence. <br> Number of bounding boxes per class: 1,828,978 cells, 134,089 columns and 316,901 rows. The layout model of Document Intelligence was used with 2024-02-29-preview API version.

Evaluation Results

The primary evaluation set is a cut of the Azure labels and digital corpora images. Number of bounding boxes per class: 200,840 cells, 13,670 columns and 34,575 rows. Mean Average Precision (mAP) was used as an evaluation metric, which measures the model's ability to correctly identify and localize objects across different confidence thresholds.

Data collection method by dataset: Hybrid: Automated, Human <br> Labeling method by dataset: Hybrid: Automated, Human <br> Properties: We evaluated with Azure labels from manually selected pages, as well as manual inspection on public PDFs and powerpoint slides.

Per-class Performance Metrics: | Class | AP (%) | AR (%) | |:-------|:-------|:-------| | cell | 58.365 | 60.647 | | row | 76.992 | 81.115 | | column | 85.293 | 87.434 |

<!---

Inference:

Acceleartion Engine: TensorRT <br> Test hardware: See Support Matrix from NIM documentation --->

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. <br> For more detailed information on ethical considerations for this model, please see the Explainability, Bias, Safety & Security, and Privacy sections below. <br> Please report security vulnerabilities or NVIDIA AI Concerns here.

Bias

Explainability

Privacy

Safety