About adapting the map crop size - Githubissues

facebookresearch / OrienterNet

Source Code for Paper "OrienterNet Visual Localization in 2D Public Maps with Neural Matching"

Other

463 stars 48 forks source link

About adapting the map crop size #25

Closed Zhenghao97 closed 8 months ago

Zhenghao97 commented 1 year ago

Hello Sarlin,

The city's content which I use is very sparse such like buildings, etc. I have to enlarge the map size to contain more content to assist localization. So If I adapt map crop size, like 64 to 96, which other params also need to adapt?

Looking forward your reply, thanks!

sarlinpe commented 1 year ago

Given https://github.com/facebookresearch/OrienterNet/blob/5f0d2d4ac27daa4319f9a6b30cec434303aa1a35/maploc/conf/data/mapillary.yaml#L30-L31

https://github.com/facebookresearch/OrienterNet/blob/5f0d2d4ac27daa4319f9a6b30cec434303aa1a35/maploc/conf/orienternet.yaml#L10-L11

You want crop_size_meters = 96
You can update max_init_error ~ crop_size_meters - z_max/2 = 80 to make the training more difficult.
If objects are sparse, you could increase z_max to 48.

All of this will increase the GPU memory requirement. A trade-off between size and accuracy is to decrease pixel_per_meter, e.g. to 1?