search

facebookresearch / OrienterNet

Source Code for Paper "OrienterNet Visual Localization in 2D Public Maps with Neural Matching"
Other
463 stars 48 forks source link

actual scale for the corresponding BEV #16

Open yejy53 opened 1 year ago

yejy53 commented 1 year ago

I have a question, what is the actual scale for the corresponding BEV? For example, if I use the demo, the generated BEV square is 64*129, and the actual scale for each grid is (m)?

yejy53 commented 1 year ago

Suppose I have a panoramic image now, how should I modify the generation of BEV, so that it can no longer be in the frustum space

sarlinpe commented 1 year ago

Sorry for the late reply. The resolution is 2 pixels per meter: https://github.com/facebookresearch/OrienterNet/blob/71e89ebdfa489e8c3d28ec668d9dad7da2ca5988/maploc/conf/orienternet.yaml#L12

https://github.com/facebookresearch/OrienterNet/blob/71e89ebdfa489e8c3d28ec668d9dad7da2ca5988/maploc/conf/data/mapillary.yaml#L29

This can be changed during data processing.

martin-liao commented 9 months ago

Sorry for the late reply. The resolution is 2 pixels per meter:

https://github.com/facebookresearch/OrienterNet/blob/71e89ebdfa489e8c3d28ec668d9dad7da2ca5988/maploc/conf/orienternet.yaml#L12

https://github.com/facebookresearch/OrienterNet/blob/71e89ebdfa489e8c3d28ec668d9dad7da2ca5988/maploc/conf/data/mapillary.yaml#L29

This can be changed during data processing.

The paper mentioned BEV has size LXD=32x32m with resolution delta=50cm, which means the shape of bev features should be 64x64 equally. However, in link, the grid_size is set to [129 (64 * 2 + 1,64). Maybe the additional 1 is a dummy bins like issue30?

sarlinpe commented 9 months ago

Not a dummy bin - we chose to position the camera at the center of the middle cell so the width of the BEV should be odd. The paper is incorrect, the size of the BEV is actually LxD=65x32m - I will try to fix it at some point.

martin-liao commented 9 months ago

Oh, I understand. Thank you!