This is the official PyTorch implementation of the paper SDNet: Spatial Adversarial Perturbation Local Descriptor Learned with the Dynamic Probabilistic Weighting Loss

Maybe some extra modules are needed. I'm not sure.

• Python 3.11
• PyTorch 2.1
• OpenCV-Python 4.8.1
• Scipy

Both the PhotoTourim datasets and HPatches datasets will be automatically downloaded when the training / testing script runs. By default they will be downloaded to ./data/. We use *dot * to separate the dataset name and the split name. PhotoTourism(Brown) datasets are named as "brown.liberty", "brown.yosemite" and "brown.notredame" respectively. We use the full HPatches dataset during training, which is named as "HP.full".

Use the following command to train a hynet-based model with the liberty sequence of PhotoTourism.

train_sp.py --num_epoch 200 --test_every_epoch --save_every_epoch --bs 1024 --optim adam --lr 0.001 --lr_policy cos --arch_type hynet --drop_rate 0.3 --out_dim 128 --loss_type sdnet --margin 0.45 --alpha_init 0 --alpha_moment 0.995 --quantile 0.0 1.0 --lambda_clean 1 --lambda_adv 1 --adv_step 0.1 --adv_iter 3 --train_data brown.liberty --test_data brown.yosemite brown.notredame HP.a --patch_size 32

For hardnet-based model:

CUDA_VISIBLE_DEVICES=0 python train_sp.py --num_epoch 200 --test_every_epoch --save_every_epoch --bs 1024 --optim adam --lr 0.001 --lr_policy cos --arch_type l2net --drop_rate 0.1 --out_dim 128 --loss_type sdnet --margin 0.45 --alpha_init 0 --alpha_moment 0.995 --quantile 0.0 1.0 --lambda_clean 1 --lambda_adv 1 --adv_step 0.1 --adv_iter 3 --train_data brown.liberty --test_data brown.yosemite brown.notredame HP.a --patch_size 32

Use the following command to test a real - valued descriptor model:

CUDA_VISIBLE_DEVICES=0 python test.py --model_dir ./pretrained/SDNet_LIB.state_dict --test_data brown.liberty brown.yosemite brown.notredame HP.a --arch_type hynet

This code borrows heavily from dynamic-soft-margin-pytorch. We thank Zhang, Linguang and Rusinkiewicz, Szymon for releasing their code.