WORK IN PROGRESS. This is an image segmentation model trained on the COCO Stuff 2017 dataset. I used the validation set because it's smaller. You can use the full train set (18 GB) because the dataset is streamed during training rather than being loaded all at once, it'll just take longer to finish an epoch.

• Train set images (18GB): http://images.cocodataset.org/zips/train2017.zip
• Val set images (1GB): http://images.cocodataset.org/zips/val2017.zip
• Image masks (both train and val): http://images.cocodataset.org/annotations/annotations_trainval2017.zip

Model parameters

• Residual U-Net (8,115,831 parameters)
• Depth: 4
• Filters: 16
• Images resized to 256x256
• Loss: Dice + Crossentropy
• Optimizer: Adam