Skip to content

Latest commit

 

History

History

mcibiplusplus

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 

README.md

Introduction

Official Repo

Code Snippet

MCIBI++ (TPAMI'2022) 
@article{jin2022mcibi++,
    title={MCIBI++: Soft Mining Contextual Information Beyond Image for Semantic Segmentation},
    author={Jin, Zhenchao and Yu, Dongdong and Yuan, Zehuan and Yu, Lequan},
    journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
    year={2022},
    publisher={IEEE}
}

Results of Different Frameworks

LIP

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU Download
FCN ImageNet-1k-224x224 R-50-D8 473x473 LR/POLICY/BS/EPOCH: 0.01/poly/32/150 train/val 51.13% cfg | model | log
PSNet ImageNet-1k-224x224 R-50-D8 473x473 LR/POLICY/BS/EPOCH: 0.01/poly/32/150 train/val 52.93% cfg | model | log
UperNet ImageNet-1k-224x224 R-50-D8 473x473 LR/POLICY/BS/EPOCH: 0.01/poly/32/150 train/val 53.92% cfg | model | log
DeepLabV3 ImageNet-1k-224x224 R-50-D8 473x473 LR/POLICY/BS/EPOCH: 0.01/poly/32/150 train/val 53.59% cfg | model | log

ADE20k

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU Download
FCN ImageNet-1k-224x224 R-50-D8 512x512 LR/POLICY/BS/EPOCH: 0.01/poly/16/130 train/val 43.39% cfg | model | log
PSNet ImageNet-1k-224x224 R-50-D8 512x512 LR/POLICY/BS/EPOCH: 0.01/poly/16/130 train/val 43.88% cfg | model | log
UperNet ImageNet-1k-224x224 R-50-D8 512x512 LR/POLICY/BS/EPOCH: 0.01/poly/16/130 train/val 44.30% cfg | model | log
DeepLabV3 ImageNet-1k-224x224 R-50-D8 512x512 LR/POLICY/BS/EPOCH: 0.01/poly/16/130 train/val 44.85% cfg | model | log

CityScapes

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU Download
FCN ImageNet-1k-224x224 R-50-D8 512x1024 LR/POLICY/BS/EPOCH: 0.01/poly/8/220 train/val 78.77% cfg | model | log
PSNet ImageNet-1k-224x224 R-50-D8 512x1024 LR/POLICY/BS/EPOCH: 0.01/poly/8/220 train/val 79.91% cfg | model | log
UperNet ImageNet-1k-224x224 R-50-D8 512x1024 LR/POLICY/BS/EPOCH: 0.01/poly/8/220 train/val 80.05% cfg | model | log
DeepLabV3 ImageNet-1k-224x224 R-50-D8 512x1024 LR/POLICY/BS/EPOCH: 0.01/poly/8/220 train/val 80.72% cfg | model | log

COCOStuff-10k

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU Download
FCN ImageNet-1k-224x224 R-50-D8 512x512 LR/POLICY/BS/EPOCH: 0.001/poly/16/110 train/test 37.38% cfg | model | log
PSNet ImageNet-1k-224x224 R-50-D8 512x512 LR/POLICY/BS/EPOCH: 0.001/poly/16/110 train/test 38.47% cfg | model | log
UperNet ImageNet-1k-224x224 R-50-D8 512x512 LR/POLICY/BS/EPOCH: 0.001/poly/16/110 train/test 39.20% cfg | model | log
DeepLabV3 ImageNet-1k-224x224 R-50-D8 512x512 LR/POLICY/BS/EPOCH: 0.001/poly/16/110 train/test 38.94% cfg | model | log

SOTA Results

VSPW

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU Download
UperNet ImageNet-1k-224x224 R-101-D8 512x512 LR/POLICY/BS/EPOCH: 0.01/poly/16/240 train/val 43.21% cfg | model | log
UperNet ImageNet-22k-384x384 Swin-Large 512x512 LR/POLICY/BS/EPOCH: 0.00006/poly/16/240 train/val 56.04% cfg | model | log

PASCAL-VOC

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU Download
UperNet ImageNet-1k-224x224 R-50-D8 512x512 LR/POLICY/BS/EPOCH: 0.01/poly/16/60 train/val 79.48 cfg | model | log
UperNet ImageNet-1k-224x224 R-101-D8 512x512 LR/POLICY/BS/EPOCH: 0.01/poly/16/60 train/val 80.42 cfg | model | log

PASCAL-Context-59

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU/mIoU (ms+flip) Download
UperNet ImageNet-1k-224x224 R-101-D8 480x480 LR/POLICY/BS/EPOCH: 0.004/poly/16/260 train/val 55.63%/56.82% cfg | model | log
UperNet ImageNet-1k-224x224 S-101-D8 480x480 LR/POLICY/BS/EPOCH: 0.004/poly/16/260 train/val 56.83%/57.92% cfg | model | log
UperNet ImageNet-22k-384x384 Swin-Large 480x480 LR/POLICY/BS/EPOCH: 0.00006/poly/16/260 train/val 62.37%/64.01% cfg | model | log

LIP

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU/mIoU (flip)/mIoU (ms+flip) Download
UperNet ImageNet-1k-224x224 R-101-D8 473x473 LR/POLICY/BS/EPOCH: 0.01/poly/32/150 train/val 55.87%/56.26%/56.32% cfg | model | log
UperNet ImageNet-1k-224x224 S-101-D8 473x473 LR/POLICY/BS/EPOCH: 0.01/poly/32/150 train/val 56.57%/56.77%/57.08% cfg | model | log
UperNet ImageNet-22k-384x384 Swin-Large 473x473 LR/POLICY/BS/EPOCH: 0.00006/poly/16/110 train/val 59.58%/59.89%/59.91% cfg | model | log
DeepLabV3 ImageNet-1k-224x224 HRNetV2p-W48 473x473 LR/POLICY/BS/EPOCH: 0.007/poly/40/150 train/val 56.70%/57.27%/57.42% cfg | model | log

ADE20k

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU/mIoU (ms+flip) Download
UperNet ImageNet-1k-224x224 R-101-D8 512x512 LR/POLICY/BS/EPOCH: 0.01/poly/16/130 train/val 46.38%/47.93% cfg | model | log
UperNet ImageNet-1k-224x224 S-101-D8 512x512 LR/POLICY/BS/EPOCH: 0.004/poly/16/180 train/val 47.59%/48.56% cfg | model | log
UperNet ImageNet-22k-384x384 Swin-Large 640x640 LR/POLICY/BS/EPOCH: 0.00006/poly/16/130 train/val 53.48%/54.50% cfg | model | log

CityScapes

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU (ms+flip) Download
DeepLabV3 ImageNet-1k-224x224 R-101-D8 512x1024 LR/POLICY/BS/EPOCH: 0.01/poly/16/440 trainval/test 82.20% cfg | model | log
DeepLabV3 ImageNet-1k-224x224 S-101-D8 512x1024 LR/POLICY/BS/EPOCH: 0.004/poly/16/440 trainval/test 81.70% cfg | model | log
DeepLabV3 ImageNet-1k-224x224 HRNetV2p-W48 512x1024 LR/POLICY/BS/EPOCH: 0.01/poly/16/440 trainval/test 82.74% cfg | model | log

COCOStuff-10k

Segmentor Pretrain Backbone Crop Size Schedule Train/Eval Set mIoU/mIoU (ms+flip) Download
UperNet ImageNet-1k-224x224 R-101-D8 512x512 LR/POLICY/BS/EPOCH: 0.001/poly/16/110 train/test 40.41%/41.84% cfg | model | log
UperNet ImageNet-1k-224x224 S-101-D8 512x512 LR/POLICY/BS/EPOCH: 0.001/poly/32/150 train/test 41.81%/42.71% cfg | model | log
UperNet ImageNet-22k-384x384 Swin-Large 512x512 LR/POLICY/BS/EPOCH: 0.00006/poly/16/150 train/test 49.11%/50.27% cfg | model | log

More

You can also download the model weights from following sources:

In addition, in this repo, all of models above are evaluated on A100 rather than V100 mentioned in our original paper, thus the performance here will be slightly different from the reported results in the original paper.