This is an extesion that adds support for the homography transformation for the model proposed in "End-to-end weakly-supervised semantic alignment" by I. Rocco, R. Arandjelović and J. Sivic.
Credit for the authors. For more information check out the original project [website] and the paper on [arXiv].
A dev docker container is provided for easy reproduction. To use it open the project in VSCode and open the project in a dev-container using the Dockerfile located at the root of the project.
The code includes scripts for pre-training the homography model with strong supervision (train_strong_hom.py) as well as to fine-tune the model using weak supervision (train_weak_hom.py).
Training scripts can be found in the scripts/ folder.
Evaluation is implemented in the eval_hom.py file. It can evaluate a homography model (with the --model-hom parameters). No evaluation for combined models is provided
The evaluation dataset is passed with the --eval-dataset parameter.
Trained models for the baseline method using only strong supervision and the proposed method using additional weak supervision are provided below. You can store them in the trained_models/ folder.
CNNGeometric with ResNet-101 baseline: [hom model]
python eval_hom.py --feature-extraction-cnn resnet101 --model-hom trained_models/cnngeo_resnet101_hom.pth.tar --eval-dataset pf-pascal
Proposed method: [hom model]
python eval_hom.py --feature-extraction-cnn resnet101 --model-hom trained_models/weakalign_resnet101_hom.pth.tar --eval-dataset pf-pascal
If you want a more in depth view of the method check the original work.