The pre-trained model checkpoint with backbone Xception65 was used for for images and MobileNetV2 for videos (for to faster inference). The pre-trained checkpoints can be obtained from https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md
The result on a single image:
