'Rethinking Atrous Convolution for Semantic Image Segmentation' 1. Pre-traind Models deeplabv3-voc2012.pt - Trained on VOC 2012 training set for 54,000 steps.- Average mIoU on VOC 2012 validation set: 0.6563 2. Prediction on VOC 2012 Validation Set Ground truth Prediction Ground truth Prediction