Question about BEV generation

[yejy53]

How to generate BEV features using panoramic data, such as downloaded Google panoramic images?
Additionally, for example, if I have panoramic imagery, can I divide it into 4 single-view images, and the geographical location of the four views is consistent, which can improve the positioning accuracy?

[sarlinpe]

You can generate perspective views from the panoramic image, infer a heatmap for each of them, and fuse the heatmaps into a single one using the relative poses between the cameras. Since perspective views are related by a rotation only, you only need to shift each heatmap along the rotation dimension by yaw_relative / 360 * num_rotations.

[yejy53]

I've taken this issue into account, but during the final implementation, when multiple BEVs are combined, stitching lines become evident. This is because BEV exhibits anomalies at the frustum edges.