Training new models

First make sure that the segmappy package is installed (See Install SegmapPy python package).

Train CNN model

To train the data-driven model:

$ source ~/segmappyenv/bin/activate
(segmappyenv)$ segmappy_train_cnn

Once the training has ended, the semantic layers should also be trained:

(segmappyenv)$ segmappy_train_semantics

Test CNN model

You can plot the ROC for the SegMap descriptor in comparison with the eigenvalue based features.

(segmappyenv)$ segmappy_plot_roc_from_matches

You can also plot how quickly descriptors extracted from incrementally grown segments contain relevant information that can be used for localization. The x-axis represents the growing status of a segment until all its measurements have been accumulated. The log-scaled y-axis represents how many neighbours in the target map need to be considered in order to include the correct target segment (the lower the better).

(segmappyenv)$ segmappy_plot_acc_versus_size

The plots will be saved as roc.pdf and acc_versus_size.pdf respectively.