Delving-Deeper-Into-Anti-Aliasing-in-ConvNets

This work is accepted in BMVC2020 as Best Paper Award. It introduces a plugin module in neural network to improve both model accuracy and consistency.

[Project page] | [arXiv] | [Slide] | [Video] | [视频]

Progress

• Image Classification
• Instance Segmentation
• Semantic Segmentation

lastest: I have released beta version Instance Segmentation code for COCO. Please refer to maskrcnn branch

Installation

torch==1.1.0
torchvision==0.2.0

Dataset

• Download the ImageNet dataset and move validation images to labeled subfolders
  • To do this, you can use the following script: https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh

File Structure

anti-aliasing
└── data
    ├── output
    ├── ILSVRC2012
└── master
    └── Adaptive-anti-Aliasing
        └── ...

Model Zoo

Model Name		Top-1 Acc	Consistency
resnet101_k3_pasa_group8_softmax	weight	79.0	91.8
resnet101_k5_pasa_group8_softmax	weight	78.6	92.2

Testing

python main.py --data ../../data/ILSVRC2012 -f 3 -e -b 32 -a resnet101_pasa_group_softmax --group 8 --weights /pth/to/model

Training

python main.py --data ../../data/ILSVRC2012 -f 3 -b 128 -ba 2 -a resnet101_pasa_group_softmax --group 8 --out-dir /pth/to/output/dir

Instance Segmentation and Semantic Segmentation

Please directly put "Adaptive-anti-Aliasing/models_lpf/layers/pasa.py" this module before downsampling layers of the backbone except the first convolution layer. We adopt implemantation directly from:

Instance Segmentation: MaskRcnn

Semantic Segmentation: Deeplab V3+ and TDNet

Citation

@inproceedings{zou2020delving,
  title={Delving Deeper into Anti-aliasing in ConvNets},
  author={Xueyan Zou and Fanyi Xiao and Zhiding Yu and Yong Jae Lee},
  booktitle={BMVC},
  year={2020}
}

Acknowledgement

We borrow most of the code from Richard Zhang's Repo (https://github.com/adobe/antialiased-cnns) Thank you : )