Distortion Aware Radial Swin Transformer

This repo is the official implementation of "DarSwin : Distortion Aware Radial Swin Transformer". It currently includes code and models for the following tasks:

Image Classification: Included in this repo. See get_started.md for a quick start.

Introduction

Distortion Aware Radial Swin Transformer (DarSwin) is initially described in arxiv, which is a backbone based on Swin Transformer of distortion aware network for wide-angle image classification. The distortion-aware radial patches enable a better generalization to unseen lenses.

Main Results on ImageNet2010

Top-1 classification accuracy (mean) as a function of test distortion for our method (DarSwin-A) and previous state of the art: Deformable Attention Transformer (DAT), Swin Transformer, and Swin Transformer + undistortion (see text). All methods are trained on a restricted set of lens distortion curves (indicated by the pink shaded regions): (a) Very low, (b) low, (c) medium and (d) high distortion. We observe zero-shot adaptation to lens distortion of each approach by testing across all $\xi \in [0, 1]$.

Citing DarSwin Transformer

@article{athwale2023darswin,
    title={DarSwin : Distortion Aware Radial Swin Transformer},
    author={Athwale, Akshaya and Afrasiyabi, Arman and Lagüe, Justin and Shili, Ichrak and Ahmad, Ola and Lalonde, Jean-Fran{\c{c}}ois},
    journal={IEEE/CVF International Conference on Computer Vision (ICCV)},
    year={2023}
  }  

Getting Started

• For Image Classification, please see get_started.md for detailed instructions.