How To Convert COCO JSON to Tensorflow TFRecord



Convert to

Tensorflow TFRecord

COCO is a common JSON format used for machine learning because the dataset it was introduced with has become a common benchmark.





Tensorflow TFRecord

Tensorflow TFRecords are a binary format used with the TensorFlow Object Detection models.

How To Convert COCO JSON to Tensorflow TFRecord

If your labeling tool exported annotations in the

COCO

format, but you’re trying to use a

MobileNet SSD v2

model that needs

TFRecord

annotations, we’ve got you covered. You can convert those in 3 clicks with Roboflow.

Step 1: Create a Free Roboflow Public Workspace

Roboflow is the universal conversion tool for computer vision annotation formats. The Public plan is the best way for those exploring personal projects, class assignments, and other experiments to try Roboflow. To convert your dataset, start by creating a workspace on the Public plan.

Create Roboflow Workspace

Step 2: Upload your data into Roboflow

Once your account has been created, click Create Dataset.

Upload your data to Roboflow by dragging and dropping your

COCO JSON

images and annotations into the upload space.

To learn how to create COCO JSON yourself from scratch, see our CVAT (object detection annotation tool) tutorial.

Step 3: Generate Dataset Version

Next, you can choose Preprocessing and Augmentation options for your dataset version and then click Generate.

Step 4: Export Dataset Version

After generating, you will be prompted to Export your dataset. You can choose to receive your dataset as a .zip file or a curl download link. You can also choose which format you would like to export. Choose

Tensorflow TFRecord

Congratulations, you have successfully converted your dataset from

COCO JSON

format to

Tensorflow TFRecord

format!

Next Steps

Ready to use your new

TFRecord

dataset?

If you're looking for more info, we have a more detailed blog post about creating Object Detection TFRecord files.

Here are some compatible models:

MobileNet SSD v2

This architecture provides good realtime results on limited compute. It's designed to run in realtime (30 frames per second) even on mobile devices.

Faster R-CNN

One of the most accurate object detection algorithms but requires a lot of power at inference time. A good choice if you can do processing asynchronously on a server.

EfficientDet Tensorflow 2

A scalable, state of the art object detection model, implemented here within the TensorFlow 2 Object Detection API.