Some questions about pairwise_t_matrix.

[fengxueguiren]

Hi, how should we understand these scaling operations for pairwise_t_matrix? Thanks very much.

        pairwise_t_matrix[...,0,1] = pairwise_t_matrix[...,0,1] * H / W
        pairwise_t_matrix[...,1,0] = pairwise_t_matrix[...,1,0] * W / H
        pairwise_t_matrix[...,0,2] = pairwise_t_matrix[...,0,2] / (self.downsample_rate * self.discrete_ratio * W) * 2
        pairwise_t_matrix[...,1,2] = pairwise_t_matrix[...,1,2] / (self.downsample_rate * self.discrete_ratio * H) * 2

[yifanlu0227]

Sorry for the late reply, you need to dive deep a little bit to the affine_grid and grid_sample function, which treat the BEV feature map as an image wrapping transformation. Here is an illustration

[Xutian-Ming]

Hi, could you please tell me where this screenshot comes from, I want to learn more details about this topic, such as why the Aj->i matrix should be like in this form and why this method can work well for multi agents collaboration perception. I wonder if you could please give more guidance.
Thank you very much!

[Xutian-Ming]

Sorry for the late reply, you need to dive deep a little bit to the affine_grid and grid_sample function, which treat the BEV feature map as an image wrapping transformation. Here is an illustration

sorry for forgetting to quote (: