The Tensorflow Object Detection API uses a proprietary binary file format called TFRecord. Creating TFRecord files has long been the bane of many developers' existence. While they are very efficient for TensorFlow's deep learning framework to parse, they are quite opaque and are not human readable. They're quite finicky; If something is wrong, it's not easy to find out why.
It's not really possible to create these files by hand. You'll need to use a tool to create them. The most foolproof tool to use is Roboflow because, unlike most one-off shell scripts, Roboflow is a universal converter for computer vision formats.
With Roboflow you can seamlessly convert whatever filetype your annotation tool exports directly into a TFRecord that's ready to use for training your model.
With Roboflow, you can deploy a computer vision model without having to build your own infrastructure.
MobileNet SSD v2
EfficientDet (D7) Tensorflow 2
models all use the
This architecture provides good realtime results on limited compute. It's designed to run in realtime (30 frames per second) even on mobile devices.
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.