This repository is a tensorflow implementation for Y. Ono, E. Trulls, P. Fua, K.M. Yi, "LF-Net: Learning Local Features from Images". If you use this code in your research, please cite the paper.
Do NOT use the ratio test for descriptor matching! The commonly-used ratio test depends on the distribution of descriptor distances, and the threshold differs from one descriptor to another. Commonly used thresholds (0.9 0.7) are actually harmful for LF-Net. If you want to use the ratio test, you need to either tune this manually, or use statistical analysis as Lowe did for SIFT.
This code is based on Python3 and tensorflow with CUDA-9.0. For more details on
the required libraries, see requirements.txt. You can also easily prepare
this by doing
pip install -r requirements.txt
Download the pretrained
models and
the scare_coeur
sequence. Extract
them to the current folder so that they fall under release/models/outdoor for
example.
For other datasets, we do not plan to release them at the moment. Please do not contact us for explanations on the training phase. We are providing them as is as a reference implementation.
The provided pre-trained models are trained with full 360 degree augmentation for orientation. Thus, the results you get from these models are slightly different from the one reported in arXiv. We have further included a consistency term on the orientation assignment.
To run LF-Net for all images in a given directory, simply type:
python run_lfnet.py --in_dir=images --out_dir=outputs
In addition, you can easily do the 2-view matching demo through
notebooks/demo.ipynb .
Training code can be found in train_lfnet.py. We will not provide any
support for the training process and datasets. All issues related to this topic
will be closed without answers.
| Outdoor dataset Top: LF-Net, Bottom: SIFT |
Indoor dataset Top: LF-Net, Bottom: SIFT |
Webcam dataset Top: LF-Net, Bottom: SIFT |
|---|---|---|
 |
 |
 |