Skip to content
Implementation of the PoseNet Architecture in Keras
Branch: master
Clone or download
Latest commit 967ed6d Dec 19, 2016
Type Name Latest commit message Commit time
Failed to load latest commit information. Initial Commit Dec 19, 2016 Initial Commit Dec 19, 2016 Initial Commit Dec 19, 2016


This is an implementation for Keras of the PoseNet architecture

As described in the ICCV 2015 paper PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization Alex Kendall, Matthew Grimes and Roberto Cipolla []

Note that this requires the TensorFlow backend for Keras, however, only minor modifications to the model as well as the file would be required in order to use the Theano backend. If someone would like assistance with this, simply open an issue.

Getting Started

  • Download the Cambridge Landmarks King's College dataset from here.

  • Download the starting and trained weights from here.

  • The PoseNet model is defined in the file

  • The starting and trained weights (posenet.npy and trained_weights.h5 respectively) for training were obtained by converting caffemodel weights from here and then training.

  • To run:

    • Extract the King's College dataset to wherever you prefer
    • Extract the starting and trained weights to the same location as and
    • Update the dataset path on line 9 in
    • If you want to retrain, simply run (note this will take a long time)
    • If you just want to test, simply run
You can’t perform that action at this time.