The extension part is based on Mask2Former codebase and please refer to Mask2Former for implementation details. Note that you might need to install detectron2 following the Mask2Former repo and implement WSIS task. Also, make a soft link to avoid repetitive download of datasets.
The generated WSIS dataset are listed below.
Dataset | MIST PGT |
---|---|
VOC2012 | trainaug set |
COCO2017 | train set |
# Performing WSIS on VOC12 dataset.
python train_net_maskrcnn.py --num-gpus 4 \
--config-file ./Mask2Former/configs/voc12/instance-segmentation/mask_rcnn_R_50_FPN_1x.yaml --dist-url tcp://0.0.0.0:12425 \
OUTPUT_DIR ./Mask2Former/results/mrcnn_voc
python train_net.py --num-gpus 4 \
--config-file ./Mask2Former/configs/voc12/instance-segmentation/maskformer2_R50_bs16_50ep.yaml --dist-url tcp://0.0.0.0:12425 \
OUTPUT_DIR ./Mask2Former/results/m2f_voc
# Performing WSIS on coco dataset.
python train_net_maskrcnn.py --num-gpus 4 \
--config-file ./Mask2Former/configs/coco/instance-segmentation/mask_rcnn_R_50_FPN_1x_wsss.yaml \
OUTPUT_DIR ./Mask2Former/results/mrcnn_coco
python train_net.py --num-gpus 4 \
--config-file ./Mask2Former/configs/coco/instance-segmentation/maskformer2_R50_bs16_50ep_wsss.yaml --dist-url tcp://0.0.0.0:12425 \
OUTPUT_DIR ./Mask2Former/results/m2f_coco