rotation-net

adaption of HomographyNet (regression head) by DeTone et al. for learning rotations of images with pytorch/tensorflow

Content of repo

Image Pairs Generation

01_data_generation.ipynb cut out grayscale patch A from MS COCO image and randomly rotate patch A between -180° and 180° to get patch B

Training

02_network_training_torch.ipynb / 02_network_training_tf.ipynb train and evaluate HomographyNet with train and val image pairs

Test

03_network_inference_torch.ipynb / 03_network_inference_tf.ipynb test with test pairs and calculate rotation error

• mean rotation error: 4.1°
• median rotation error: 2.5°

Prerequisites

• download MS COCO dataset (train/val/test) and unzip to data/ folder
• (optional) download trained model here

Credits

• mazenmel/Deep-homography-estimation-Pytorch (data generation, pytorch implementation)
• richard-guinto/homographynet (tensorflow implementation)