Keras RMAC

Regional Maximum Activations of Convolutions (RMAC) feature extractor for Keras, based on (Tolias et al. 2016) and (Gordo et al. 2016). The architecture of the model is as in the image below:

RoiPooling code from: https://github.com/yhenon/keras-spp

Prerequisites

This code requires Keras version 2.0 or greater.

• Python (2.7)
• Keras (2.1.2)
• Theano (0.9.0)
• VGG16 weights --> download the file and save it in data/ folder

References

• Tolias, G., Sicre, R., & Jégou, H. Particular object retrieval with integral max-pooling of CNN activations. ICLR 2016.

• Gordo, A., Almazán, J., Revaud, J., & Larlus, D. Deep image retrieval: Learning global representations for image search. ECCV 2016.