Neural Feature Filtering for Faster Structure from Motion Localization

The code for Neural Feature Filtering for Faster Structure from Motion Localization

The files themselves will contain comments on how to use them.

To get data for Match and No Match, run the following commands:

• create_universal_models.py
• get_points_3D_mean_desc_ml_mnm.py
• create_training_data_and_train_for_match_no_match.py

To get data for Predicticting Matchability, run the following commands:

• create_universal_models.py (no need to run again if you have already run it, from above)
• create_training_data_predicting_matchability.py
• train_for_predicting_matchability.py

To get data for Neural Filtering, run the following commands:

• create_nf_training_data.py
• train_for_nf.py

Benchmarking the models and generating statistics:

• learned_models_benchmarks.py #will generate model statistics (this will be merged with the next script)
• learned_models_pose_data.py #will generate pose data statistics
• parse_results_for_thesis.py #this file might need manualy adjustments

Notes:

In most scripts you have to pass, 'HGE' or 'CMU' or 'RetailShop' to specify the dataset