Implementation to the L-STN network in our ISASD, 2024 paper "An Improved Spatial Transformer Network based on Lightweight Localization Net (L-STN)"
Spatial transformer network (STN) is a powerful module that improves the spatial invariance of convolutional neural networks. Amid all components, the localization net serves as the backbone as it intakes the feature map and generates the affine transformation parameters, deciding how the transformation to be operated. Meanwhile, the design of the localization net is closely correlated to the outcome of transformation consequently affecting the system performance. This work simulates through different localization net designs and examines them with several prominent models on various datasets, and which aims to discover a modern localization network architecture that boosts the performance of STN transformation with less parameters and computational overhead.
code/STN
├── Caltech101
│ ├── alexnet_sp.ipynb
│ ├── DenseNet121_sp.ipynb
│ ├── DenseNet161_sp.ipynb
│ ├── DenseNet169_sp.ipynb
│ ├── DenseNet201_sp.ipynb
│ ├── EfficientNetb0_sp.ipynb
│ ├── EfficientNetb1_sp.ipynb
│ ├── EfficientNetb2_sp.ipynb
│ ├── EfficientNetb3_sp.ipynb
│ ├── EfficientNetb4_sp.ipynb
│ ├── EfficientNetb5_sp.ipynb
│ ├── ResNet-101_sp.ipynb
│ ├── ResNet-152_sp.ipynb
│ ├── ResNet-18_sp.ipynb
│ ├── ResNet34_sp.ipynb
│ └── ResNet-50_sp.ipynb
├── Cifar10
│ ├── alexnet_sp.ipynb
│ ├── DenseNet121_sp.ipynb
│ ├── DenseNet161_sp.ipynb
│ ├── DenseNet169_sp.ipynb
│ ├── DenseNet201_sp.ipynb
│ ├── EfficientNetb0_sp.ipynb
│ ├── EfficientNetb1_sp.ipynb
│ ├── EfficientNetb2_sp.ipynb
│ ├── EfficientNetb3_sp.ipynb
│ ├── EfficientNetb4_sp.ipynb
│ ├── ResNet-101_sp.ipynb
│ ├── ResNet-152_sp.ipynb
│ ├── ResNet-18_sp.ipynb
│ ├── ResNet34_sp.ipynb
│ └── ResNet-50_sp.ipynb
└── TSRD
├── alexnet_sp.ipynb
├── DenseNet121_sp.ipynb
├── DenseNet.ipynb
├── EfficientNetb0_sp.ipynb
├── EfficientNetb1_sp.ipynb
├── EfficientNetb2_sp.ipynb
├── EfficientNetb3_sp.ipynb
├── ResNet-101_sp.ipynb
├── ResNet-152_sp.ipynb
├── ResNet-18_sp.ipynb
├── ResNet-34_sp.ipynb
└── ResNet-50_sp.ipynb




