What 'viz_2d', 'correlation', 'pr' and 'pr_match' are referring to?

[Anonymous-so]

Hi, thanks for this excellent work!@dawnos log: grad_histogram: "tb" viz_2d: "none" correlation: "none" feature: "none" pr: "file" transform_image: "file" Could you please give me some hints on what 'viz_2d', 'correlation', 'pr' and 'pr_match' (in config/*.yaml file) are referring to?

[dawnos]

Thanks for your question!

Those parameters under 'log' control the outputs during training or testing. Each parameter can be 'file', 'tb' or 'none', which correspond to logging to file, logging to tensorboard, or disabled. For example, if 'grad_histogram' is set to be 'tb', then the gradients will be recorded and shown in tensorboard.

Here I will show you what does each parameter mean:

• grad_histogram: record the gradients of some of the network layers.
• transform_image: show the 'translated image', which is described in my paper in detail. For short, it is the decoded image from s_1 and a_2, where s_1 is the place feature from image 1 while a_2 is the appearance feature from image 2.
• viz_2d, viz_2d, correlation, and feature: these parameters are used in the experiments in a recently summited paper. For short, if enabled, they will visualize some 2d results for a low-dimension toy case. This case is intended for illustrating the network with quantity results.
• pr and pr_match: They are used to compute results for the place recognition task. 'pr' is used to compute AUC and ACC. 'pr_match' is used to prepare results for RobotCar-Seasons datasets, which is added in the new paper (not presented in the ICRA paper).