tps_stn_pytorch

tps_stn_pytorch

PyTorch implementation of Spatial Transformer Network (STN) with Thin Plate Spline (TPS).

Introduction

STN is a powerful neural network architecture proposed by DeepMind in [1]. STN achieves real spatial invariance by automatically rectify input images before they are feeded into a normal classification network. The most amazing part of STN is that it is end-to-end differential and can be directly plugged into existing network architectures (AlexNet, Resnet, etc), without any extra supervision.

Original STN paper [1] experiments on three specific transformation forms: Affine Transformation, Projective Transformation and Thin Plate Spline Transformation (TPS). Among them I think TPS is the most powerful translation because it can warp a image in arbitrary way. As shown below, I can warp my Avatar

into

TPS-STN has been used in OCR application [2]. In this paper TPS-STN is to automatically rectify distorted text images, before they are feeded into a normal OCR text recognition model:

Dependencies

• Python3
• PyTorch
• Torchvision
• Numpy
• Matplotlib
• Pillow / PIL
• imageio

I use imageio to create GIF visualization. Simply install it by pip install imageio.

Run

python mnist_train.py --model unbounded_stn --angle 90 --grid_size 4
python mnist_visualize.py --model unbounded_stn --angle 90 --grid_size 4
python mnist_make_gif.py --model unbounded_stn --angle 90 --grid_size 4

Then PNG and GIF resutls will be saved in ./image/unbounded_stn_angle60_grid4/ and ./gif/unbounded_stn_angle60_grid4/.

You can try other combinations of model architecture, mnist random rotation angle and TPS grid size. Details below.

Arguments

There are three controllable arguments: --model, --angle, --grid_size.

--model: str, required

• With no_stn, STN module is discarded and only a single CNN classifier remains.
• With bounded_stn, the output of localization network is squeezed to [-1, 1] by F.tanh, as was done in [2]
• With unbounded_stn, the output of locolizaition network is not squeezed

--angle: int, default = 60

• Samples in MNIST dataset will be rotated by random angles within [-angle, angle]

--grid_size: int, default = 4

• Use (grid_size x grid_size) control points to define Thin Plate Spline transformation

Test Accuracy Curves

Visualize TPS-STN

Results with angle = 90 are generally bad:

Results with bounded_stn are bad if grid_size <= 3:

But ok if grid_size >= 4:

Results with unbounded_stn are ok:

Of course there are always bad cases in each combination. You can download all my GIFs from Baidu Netdisk (file size 2G).

Reference

[1] Spatial Transformer Networks

[2] Robust Scene Text Recognition with Automatic Rectification

[3] 数值方法——薄板样条插值（Thin-Plate Spline）