Open Images Pre-trained Semantic Segmentation

Semantic segmentation assigns every pixel in an image to a class label. Semantic segmentation does image classification at pixel level. Unlike instance segmentation which can label individual instances belonging to a class, semantic segmentation clubs all instances of a class to same label. This model object contains pretrained weights that may be used as a starting point with the following semantic segmentation networks in TAO Toolkit to facilitate transfer learning.

The following semantic segmentation architecture are supported:

  • UNet

Supported Backbones

The pre-trained weights are trained on a subset of the Google OpenImages dataset. Following backbones are supported with these UNet networks.

  • resnet10/resnet18/resnet34/resnet50/resnet101

  • vgg16/vgg19


  • These are unpruned models with just the feature extractor weights, and may not be used without retraining to deploy in a classification application.

  • Please make sure to set the all_projections field to False in the spec file when training a ResNet101 model.

For more instructions on downloading and using the models defined here, refer to the NGC catalog page.