Based on Jan and Jason's work, this is a working repo.
Camera Pose Estimation originally based on Kalman Filter, using Deep Learning and recurrent neural networks.
Based on the work of Jason Rambach.
The program requires two input files, one holding the imu measurements and another one holding the ground truth data.
python <network_version.py> <imu_output> <ground_truth>
| Option | Short Version | Type | Meaning |
|---|---|---|---|
| --batch-size | -bs | Integer | The Batch Size |
| --display-step | -ds | Integer | How often the intermediate results should be output. Each ds-th epoch |
| --epochs | -ep | Integer | How many epochs to train |
| --output | -o | String | File to write the output to |
There is a script to plot the results called plot_results.py inside the Scripts folder.
It can plot the output files by simply calling
python plot_results.py <outputfile>
| Option | Short Version | Type | Meaning |
|---|---|---|---|
| --no-display | -nd | Flag | Does not display the plot. Useful when only the output file is desired |
| --output | -o | String | Output file where the plot should be stored.Will be stored in png format |
| --training-only | -to | Flag | Plots only the training cost |
| --validation-only | -vo | Flag | Plots only the validation cost |
| --model-output | -mo | Flag | Output file where the model should be stored |
|
This software is based on the TensorFlow Deep Learning Framework.
Unfortunately TensorFlows API has changed a lot here and there, sometimes documented often enough undocumented. That's why it may run with some version and won't with some others.
The version you will install depends on what OS and what distribution you use. It won't work with the latest 1.1 version but should work with something like 0.9.
If you have veresion 0.10.0 you can easily run the code when you change the line
init = tf.global_variables_initializer()
to
init= tf.initialize_all_varibles()
The latter is deprecated and no more available in later version, but the new version might not be available yet in an older version (like 0.10.0).
The original paper can be found here