Training data collection

lucbettaieb / thesis

Probably some kind of CNN localization stuff

2 stars 0 forks source link

Training data collection #6

Closed lucbettaieb closed 7 years ago

lucbettaieb commented 7 years ago

Bagging a bunch of images takes up a lot of processing power and space on the TK1. This can be avoided by saving images every so often within a given region.

Work started with this in the data_collector package, but needs to be addressed in a more organized way.

[ ] Determine which datasets to gather

Without bagging

[ ] Get data_collector package working
[ ] Write pertinent launch files
[ ] Write pertinent documentation
[ ] Define good heuristic for data collection (?)

lucbettaieb commented 7 years ago

Datasets to gather

Parameters

Number of regions per environment (n_regions)
Resolution of training data (resolution)
How many training images per region (n_images)

Training sets to get

1) Lab / Office Environment a. 5, 10, 20, 40 regions

2) Alternate Environment a. 5, 10, 20, 40 regions

Networks to train

Single hidden layer
Multiple hidden layers
Retraining inception