Spatial Transformer Networks

Reimplementations of:

• Spatial Transformer Networks
• Inverse Compositional Spatial Transformer Networks

Although implementations already exists, this focuses on simplicity and ease of understanding of the vision transforms and model.

Results

During training, random homography perturbations are applied to each image in the minibatch. The perturbations are composed by component transformation (rotation, translation, shear, projection), the parameters of each sampled from a uniform(-1,1) * 0.25 multiplicative factor.

Example homography perturbation:

Test set accuracy:

Sample alignment results:

Basic affine STN

Image	Samples
original perturbed transformed

Homography STN

Image	Samples
original perturbed transformed

Homography ICSTN

Image	Samples
original perturbed transformed

Mean and variance of the aligned results (cf Lin ICSTN paper)

Mean image

Image	Basic affine STN	Homography STN	Homography ICSTN
original perturbed transformed

Variance

Image	Basic affine STN	Homography STN	Homography ICSTN
original perturbed transformed

Usage

To train model:

python train.py --output_dir=[path to params.json]
                --restore_file=[path to .pt checkpoint if resuming training]
                --cuda=[cuda device id]

params.json provides training parameters and specifies which spatial transformer module to use:

1. BasicSTNModule -- affine transform localization network
2. STNModule -- homography transform localization network
3. ICSTNModule -- homography transform localization netwokr (cf Lin, ICSTN paper)

To evaluate and visualize results:

python evaluate.py --output_dir=[path to params.json]
                   --restore_file=[path to .pt checkpoint]
                   --cuda=[cuda device id]

Dependencies

• python 3.6
• pytorch 0.4
• torchvision
• tensorboardX
• numpy
• matplotlib
• tqdm

Useful resources

• https://github.com/chenhsuanlin/inverse-compositional-STN