Code for the BMVC 2016 paper Learning local feature descriptors with triplets and shallow convolutional neural networks

We provide the following pre-trained models:

TFeat has been integrated into Kornia

First install Kornia: pip install kornia

import torch
import kornia as K

input = torch.rand(16, 1, 32, 32)
tfeat = K.feature.TFeat(pretrained=True)
descs = tfeat(input) # 16x128

To run TFeat on a tensor of patches:

tfeat = tfeat_model.TNet()
net_name = 'tfeat-liberty'
models_path = 'pretrained-models'
net_name = 'tfeat-liberty'
tfeat.load_state_dict(torch.load(os.path.join(models_path,net_name ".params")))
tfeat.cuda()
tfeat.eval()

x = torch.rand(10,1,32,32).cuda()
descrs = tfeat(x)
print(descrs.size())

#torch.Size([10, 128])

Note that no normalisation is needed for the input patches, it is done internally inside the network.

We provide an ipython notebook that shows how to load and use the pre-trained networks. We also provide the following examples:

• extracting descriptors from image patches
• matching two images using openCV
• matching two images using vlfeat

For the testing example code, check tfeat-test notebook

We provide an ipython notebook with examples on how to train TFeat. Training can either use the UBC datasets Liberty, Notredame, Yosemite, the HPatches dataset, and combinations of all the datasets.

For the training code, check tfeat-train notebook