You can train object detection or semantic segmentation models directly from CVAT, which you can then use to automatically pre-annotate new data. This dramatically reduces the time it takes to train new models and iteratively improve your models. Moreover, apart from built-in models that we provide, you easily add custom model training Workflows and use them for pre-annotation as well.
Click on Open for a task you want to train a model on.
Click on Job #X, where X could be any job number. Annotate few frames. For testing you can just annotate one frame. But ideally you want to have thousands of objects to train a deep learning model on. Alternatively, you can just run pre-annotation if your labels are common ones.
Click on Actions for a task you want to train a model on. Then, click on Execute training Workflow.
Select a training Workflow Template. By default, you can use TF Object Detection Training for object detection or MaskRCNN Training for semantic segmentation.
Note you can easily add your own models as well. See our documentation for more information on adding custom models.
You can optionally copy and paste a model path from previously trained model or leave this field empty.
Click Submit. This will execute the training Workflow for the selected model. You can view real-time logs and view TensorBoard by clicking Open Workflow details.
You can also use this trained model to run automatic pre-annotation in CVAT. See documentation for more information on automatic pre-annotation.