Convolutional Autoencoder for Loop Closure 2.0.
To get started, download the COCO dataset and the "stuff" annotations, then run dataset/gen_tfrecords.py.
Make sure to unzip the tar in the dataset directory first.
Doing this will generate the sharded tfrecord files as well as loss_weights.txt.
After that you can train with calc2.py.
Check the --mode options in calc2.py to see what else you can do, like PR curves and finding the best model in a directory.
If you use this code for your research, please cite our paper:
@InProceedings{Merrill2019IROS,
Title = {{CALC2.0}: Combining Appearance, Semantic and Geometric Information for Robust and Efficient Visual Loop Closure},
Author = {Nathaniel Merrill and Guoquan Huang},
Booktitle = {2019 International Conference on Intelligent Robots and Systems (IROS)},
Year = {2019},
Address = {Macau, China},
Month = nov,
}