Use external object detection training data within ArcGIS

We designed ArcGIS deep learning tools to work seamlessly with training data in the PASCAL VOC (Pattern Analysis, Statistical Modeling and Computational Learning, Visual Object Classes) format. While we recommend creating and managing your training data directly within ArcGIS to ensure full compatibility and reliability, we understand that you may need to work with datasets originally labeled in other formats.

To help you work with your existing datasets and streamline your deep learning workflows, we’ve developed a suite of Python scripts that convert metadata from several common formats into ArcGIS-ready PASCAL VOC formats.

Why PASCAL VOC?

The PASCAL VOC format is a computer vision standard that keeps things simple. Its XML files specify bounding boxes using absolute pixel coordinates ( <xmin>, <ymin>, <xmax>, and <ymax> ) rather than percentages or ratios.

These absolute corner points differentiate PASCAL VOC from other formats. While external data varies in file type (JSON, TXT, etc.) and structure, the biggest difference lies in how they define the bounding box. Many formats record normalized values (0.0 to 1.0) or relative dimensions (center point, width, height) instead of pixels. Because ArcGIS tools require the PASCAL VOC standard, you must convert these external datasets to absolute pixel coordinates to use them successfully.

Click here to view technical specifications for key external datasets

YOLO (You Only Look Once) [x_center, y_center, width, height] where x_center and y_center represent the box center coordinates, and width and height specify the box size. All values are normalized as fractions of the image width and height. File Type: .txt
COCO (Common Objects in Context) [x, y, width, height] where x and y denote the top-left corner, and width and height specify the box size. All values are normalized as fractions of the image width and height. File Type: .json
Azure Machine Learning [x, y, width, height] where x and y denote the top-left corner, and width and height specify the box size. All values are normalized as fractions of the image width and height. File Type: .json
TFRecords [x, y, width, height] where x and y denote the top-left corner, and width and height specify the box size. All values are normalized as fractions of the image width and height. File Type: .json

Convert to PASCAL VOC

External object detection training data must be converted to the PASCAL VOC absolute pixel standard for use in ArcGIS. Our Python scripts automate this transformation.

Use the links below to download the relevant conversion script and documentation:

• YOLO to PASCAL VOC

• COCO to PASCAL VOC

• AzureML to PASCAL VOC

• TFRecords to PASCAL VOC

Work with the converted data

Once you’ve finished the conversion and have your new PASCAL VOC XML files, follow these steps to finalize your dataset and begin training:

• Replace labels—Replace your original labels folder in your dataset with the new folder containing the converted PASCAL VOC XML files.
• Verify structure—Ensure your main dataset folder (for example, trainingdata) contains only the two required subfolders: images and labels.
• Train your model—Use the Train Deep Learning Model tool and specify your main dataset folder (the one containing the images and labels subfolders) as the input.

You don’t need to start from scratch if your data isn’t in Pascal VOC. This streamlined approach helps you transform your existing datasets so they work within ArcGIS.