Our CAST can do unsupervised/fully-supervised segmentation on Pascal VOC. For unsupervised segmentation, we generate segmentations and extract segment features using CAST, which is self-supervisedly trained on COCO. For fully-supervised segmentation, we fine-tune CAST with ground-truth pixel labels on VOC. All experiments are based on SPML framework.
We provide the bashscripts for running all experiments. By default, we use MoCo-trained CAST-S.
- Unsupervised segmentation on VOC
> bash scripts/spml/nn_cast.sh
- Fine-tuning CAST on Pascal Context
> bash scripts/spml/finetune_cast.sh