This repository contains the pytorch implementation of Universal Correspondence Network, NIPS'16 (UCN) for geometric correspondences with few more improvements.

The UCN combines the metric learning (contrastive loss) with the fully-convolutional feature extraction for dense geometric and semantic correspondence.

As we are releasing this in 2019, we use a different network with pixel stride 1 for high resolution features to push the boundary of the dense features and see how far the features would go. The networks are based on Residual Networks

• U-Net crossing the encoder and decoder by default .

Similarly, we use the hardest contrastive loss proposed in the Fully Convolutional Geometric Features, ICCV'19 for fully convolutional metric learning instead of the contrastive loss. This allows using the O(~10k) points for hardest negative mining for all positives O(1k) per single image pair.

Note that the hardest contrastive loss does not require saving the intermediate features for negative mining. Once you find the hardest negatives, you can discard all intermediate computations. This is the crucial component making the metric learning of a large set of points possible.

Finally, for the ease of implementation and use, we do not use the convolutional spatial transformers (Rigid, SO(2), and Thin-Plate-Spline (TPS) predecessor of the deformable convolution) originally proposed in the UCN.

The limitations of the patch based feature learning is that: First, extracting a small image patch limits the receptive field of the network but is also computationally inefficient since all intermediate representations are computed separately even for overlapping regions. Second, this allows end-to-end training as the selection process is delegated to the next stage in the pipeline. Third, limiting feature extraction to a keypoints results in lower spatial resolution for subsequent matching stages and can thus reduce registration accuracy. However, this can also results in more computation and increase complexity in discrete keypoint selection process.

# Follow the instruction on pytorch.org to install pytorch on your environment
git clone https://github.com/chrischoy/open-ucn.git
cd open-ucn
pip install -r requirements.txt

YFCC data download and processing

bash scripts/download_yfcc.sh /path/to/download/yfcc
python -m scripts.gen_yfcc --source /path/to/download/yfcc --target /path/to/preprocessed/yfcc

Modify the arguments accordingly.

python train.py \
	--batch_size 1 \
	--iter_size 2 \
	--data_dir_raw /path/to/yfcc100m \
	--data_dir_processed /path/to/preprocessed \
	--num_hn_samples_per_batch 40000 \
	--model_n_out 96

One interesting phenomenon I found while training the network was that the training was very unstable with a smaller number of output features (i.e. feature size). For example, if I train the network to generate 64 dimensional features, the training only succeeds 1 out of 3 times. However, for 3D space, I could successfully push the dimension to 16 without a problem (link). This indicates the inherent difficulty of the 2D geometric correspondences.

Following demo code will download the UCN and test it on a few image pairs. The output will be saved on ./ucn_outputs.

Note: The code requires GPU with VRAM > 4G by default and would use the most computation heavy method for visualization. There are various NN search methods and try out different modes if things don't work.

python demo.py

To see more visualizations, visit Youtube link.

A fully-convolutional model generates dense features. Similar to SIFT, it is crucial to filter out some false positive matches with heuristic methods. Here, we use 1. reciprocity test and 2. stability test.

1. Reciprocity test simply checks whether the nearest neighbors of image 1 in image 2 fall back to the vicinity of the keypoints in image 1. The idea is that if the feature is discriminative, then it should form a cycle.

2. Stability test selects features that are shifted randomly by 1 pixel. The idea is that if the features are discriminative, it should be invariant to translation noise.

Feel free to contribute to the model zoo by submitting your weights and the architecture.

Note: The models are train only on the YFCC dataset and are not guaranteed to work on other datasets with different statistics. (e.g., white background)

Note: The models assume a gray scale image in [-0.5, 0.5], i.e. Given an uint8 image x in [0, 255], scale to (x / 255 - 0.5).

The Fully Convolutional Features for 2D Correspondences

@incollection{UCN2016,
    title = {Universal Correspondence Network},
    author = {Choy, Christopher B and Gwak, JunYoung and Savarese, Silvio and Chandraker, Manmohan},
    booktitle = {Advances in Neural Information Processing Systems 29},
    year = {2016},
}

Fully Convolutional Metric Learning and Hardest Contrastive Loss

@inproceedings{FCGF2019,
    author = {Christopher Choy and Jaesik Park and Vladlen Koltun},
    title = {Fully Convolutional Geometric Features},
    booktitle = {ICCV},
    year = {2019},
}

Open-source Pytorch Implementation

@misc{
    author = {Christopher Choy and Junha Lee},
    title = {Open Universal Correspondence Network},
    howpublished={\url{https://github.com/chrischoy/open-ucn}},
    year = {2019},
}

ASLFeat proposes fully convolutional features and keypoint detection with deformable convolution (convolutional spatial transformer), similar to Universal Correspondence Network which uses fully convolutional features with convolutional spatial transformer.

MIT License