2018-08-07 mxnet + pspnet

[yzbx]

pspnet on cityscapes, use memory 2 x 9000+ edit code in python packages: xxx/lib/pythonx.x/site-packages/glouncv/model_zoo

python models/mxnet/train.py --dataset Cityscapes --model psp --backbone resnet50 --lr 0.001 --checkname resnet50_psp_pascal --batch-size 8

Epoch 49, validation pixAcc: 0.929, mIoU: 0.552: 100%, dataset: cityscapes Epoch 49, validation pixAcc: 0.936, mIoU: 0.586: 100%, dataset: cityscapes Epoch 49, validation pixAcc: 0.949, mIoU: 0.811: 100%, dataset: voc

• TuSimple-DUC https://github.com/ISCAS007/TuSimple-DUC

  source activate env2
  python train/train_model.py configs/train/train_cityscapes.cfg

• torchseg upsample_layer = 4, default is 3

  python test/pspnet_test.py --test=naive --midnet
  _name=aspp --batch_size=4 --backbone_name=resnet50 --upsample_layer=4

mxnet + wide resnet + fcn

https://github.com/ISCAS007/ademxapp

model	training data	testing scale	class IoU (%)	class iIoU (%)	category IoU (%)	category iIoU(%)
Model A2, 2 conv.	fine	1024x2048	78.4	59.1	90.9	81.1
Model A2, 2 conv.	fine	multiple	79.4	58.0	91.0	80.1
Model A2, 2 conv.	fine; coarse	1024x2048	79.9	59.7	91.2	80.8
Model A2, 2 conv.	fine; coarse	multiple	80.6	57.8	91.0	79.1

[yzbx]

pytorch voc

python test/pspnet_test.py --test=naive --dataset_name=VOC2012 --batch_size=4 --midnet_name=aspp --backbone_pretrained=True

pytorch cityscapes coarse

python test/pspnet_test.py --test=coarse --net_name=pspnet --backbone_name=resnet50 --backbone_pretrained=True --midnet_name=aspp --augmentations_rotate=True --augmentations_blur=True --batch_size=4 --note=aug-rT-bT

python test/pspnet_test.py --test=coarse --net_name=pspnet --backbone_name=resnet50 --backbone_pretrained=True --midnet_name=aspp --augmentations_rotate=False --augmentations_blur=False --batch_size=4 --note=aug-rF-bF

semseg cityscapse

python train.py --arch=pspnet --dataset=cityscapes --batch_size=4