map-related rot question

[lsrrrrrrrrr]

Thanks for this great work. I am wonder the following reasons for the above flip and rotation in projects/mmdet3d_plugin/datasets/nuscenes_e2e_dataset.py , why these tranforms are needed:

Question 1: semantic_masks, instance_masks, forward_masks, backward_masks = preprocess_map(vectors, self.patch_size, self.canvas_size, self.map_num_classes, self.thickness, self.angle_class) instance_masks = np.rot90(instance_masks, k=-1, axes=(1, 2))

map_mask = obtain_map_info(self.nusc, self.nusc_maps, info, patch_size=self.patch_size, canvas_size=self.canvas_size, layer_names=['lane_divider', 'road_divider']) map_mask = np.flip(map_mask, axis=1) map_mask = np.rot90(map_mask, k=-1, axes=(1, 2)) map_mask = torch.tensor(map_mask.copy())

[fishmarch]

I was also puzzled with these codes. I think these are required because of mistakes in other codes. Such as, they use ego2global pose to retrieve map vectors, then they need the first np.rot90, but actually here requires lidar2global pose; there is a useless transpose in obtain_map_info, which then requires np.flip and np.rot90 to transpose back....