fast-image-retrieval

A lightweight framework using binary hash codes and deep learning for fast image retrieval.

Configuration

To run the examples, you need to create a config.cfg file under the root folder of this project. An example of config.cfg looks like:

[examples]
shoes7k_pos_path: /path/to/datasets/shoes7k/classification
shoes7k_neg_path: /path/to/datasets/shoes7k/classificationNeg

Run Examples

Shoes7k

To run example on shoes7k dataset, firstly, you need to convert shoes7k dataset to LMDB dataset.

cd fast-image-retrieval/
python ./examples/shoes7k/convert_shoes7k_data.py

Then, you need to train the CNN model.

./examples/shoes7k/train.sh

Next, you can retrieve similar image target.jpg using

./examples/shoes7k/retrieve.sh target.jpg

Note that the first retrival procedure might be very slow because the program reads all shoes7k images and generate feature matrix. The later retrivals can be very fast.

Evaluate over the whole dataset by using

python ./examples/shoes7k/eval.py

Some retrieval results: