Possible GT leakage in inference: gt_bboxes_3d/gt_labels_3d used to build prior points during testing

[RBinSky]

Thanks for releasing the code. I noticed that during testing, the dataset still provides gt_bboxes_3d and gt_labels_3d, and in MapTRv2 these GT fields are used inside make_multi_input() (called from simple_test_pts) to build the prior points. This seems to inject ground-truth map geometry into inference. Could you clarify whether this is intended? If not, what is the correct test-time setting to avoid using GT (e.g., disable GT in prepare_test_data or skip the GT branch in make_multi_input)?

[missTL]

We applied Gaussian noise to the real map as per MapEX to simulate the expired HD maps in the real world.
Since there are expired HD maps in the real world, but no corresponding expired HD maps in nuScenes, we can only follow MapEX for the noise simulation.