This repository has the source code and the Stanford Online Products dataset for the paper "Deep Metric Learning via Lifted Structured Feature Embedding" (CVPR16). The paper preprint is available on arXiv. If you just need the Caffe code, check out the Submodule. For the loss layer implementation, look at here.

If you find this work useful in your research, please consider citing:

@inproceedings{songCVPR16,
    Author = {Hyun Oh Song and Yu Xiang and Stefanie Jegelka and Silvio Savarese},
    Title = {Deep Metric Learning via Lifted Structured Feature Embedding},
    Booktitle = {Computer Vision and Pattern Recognition (CVPR)},
    Year = {2016}
}

1. Install prerequsites for Caffe (see: Caffe installation instructions)
2. Compile the Caffe-Deep-Metric-Learning-CVPR16 Github submodule.

1. Download pretrained GoogLeNet model from here
2. Download the ILSVRC12 ImageNet mean file for mean subtraction. Refer to Caffe the ImageNet examples here.
3. Generate the LMDB file to convert the training set of images to the DB format. Example scripts are in code/ directory.

• Modify and run code/compile.m to mex compile the cpp files used for LMDB generation.
• Modify code/config.m to set save paths.
• Run code/gen_caffe_dataset_multilabel_m128.m to start the LMDB generation process.

4. Create the model/train*.prototxt and model/solver*.prototxt files. Please refer to the included *.prototxt files in model/ directory for examples. You also need to provide the path to the ImageNet mean file (usually called imagenet_mean.binaryproto) you downloaded in step 2.
5. Inside the caffe submodule, launch the Caffe training procedure. caffe/build/tools/caffe train -solver [path-to-training-prototxt-file] -weights [path-to-pretrained-googlenet] -gpu [gpuid]

1. Modify and run code/gen_caffe_validation_imageset.m to convert the test images to LMDB format.
2. Modify the test set path in model/extract_googlenet*.prototxt.
3. Modify the model and test set path and run code/compute_googlenet_distance_matrix_cuda_embeddings_liftedstructsim_softmax_pair_m128.py.

1. Use code/evaluation/evaluate_clustering.m to evaluate the clustering performance.
2. Use code/evaluation/evaluate_recall.m to evaluate recall@K for image retrieval.

You can download the Stanford Online Products dataset (2.9G) from ftp://cs.stanford.edu/cs/cvgl/Stanford_Online_Products.zip

• We also have the text meta data for each product images. Please let us know if you're interested in using them.

You can download our pre-trained models on the Cars196 dataset, the CUB200 dataset and the Online Products dataset (265M) from ftp://cs.stanford.edu/cs/cvgl/pretrained_models.zip

MIT Licence