TFeat shallow convolutional patch descriptor

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

Pre-trained models

We provide the following pre-trained models:

network name	model link	training dataset
tfeat-liberty	tfeat-liberty.params	liberty (UBC)
tfeat-yosemite	tfeat-yosemite.params	yosemite (UBC)
tfeat-notredame	tfeat-notredame.params	notredame (UBC)
tfeat-ubc	coming soon...	all UBC
tfeat-hpatches	coming soon...	HPatches (split A)
tfeat-all	coming soon...	All the above

Pre-trained models usage with Kornia

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

Quick start guide

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.

Testing TFeat: Examples (WIP)

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

Re-training TFeat

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