Thank you for your work and dataset.
I was trying to use VoxFormer ckpt(trained on SemanticKITTI) to do inference on the SSC-NuScene, with different depth prediction modules, but the performance after the 1st stage was very poor compared to what's shown in the paper.
{'IoU_empty': 0.7094805990151872, 'oU_occupied': 0.13383726556742306, 'IoU': 0.13383726556742306, 'Precision': 0.7197374144135552, 'Recall': 0.1411957741150115}.

I directly deployed Lite-mono and another MVS depth prediction network on CAM_FRONT images from NuScene, and I computed 1/prediction as the depth since these network output inverse_depth. However, the depth prediction from Litemono is below around 5m thus not meaningful for generating query proposal furthermore. I also tried MVS network to predict the depth, and I visualized the voxel queries after aggregating pseudo point clouds of 9 frames in the ego coordinate, but it seems that the predicted depth is still in wrong scale. Could you please tell me how did you deal with the scale of depth in your work? And could you please tell me the connection between the id of scene/frame in your SSC-NuScene and NuScene origin dataset?

Many thanks!