About perform OCC prediction using only one camera

[cyhasuka]

Hello, first of all thank you for being able to open source this great work. I have two questions I would like to ask.

1. I currently have an RC cart with separate cameras placed in a certain direction. I would like to input only one image of the real camera (e.g. the right side) and set all the other images to black. In this case, is it possible to ensure that the input image has the correct orientation result? I know that the model may output incorrect predictions for other orientations, but I don't really care about the other orientations, as long as the orientation of the input image (e.g., the right side) OCC result turns out to be correct.
   May I ask if such a result can be achieved?

2. Since we don't have enough resources and data to re-train the model, can I directly use your pre-trained model for inference by changing the form of camera Intrinsics & Extrinsics? Will there be a significant difference in the inference results?

[AlphaPlusTT]

Thanks for your interest in our work! A few notes:

1. Our method relies heavily on LiDAR. The main performance gains come from effectively leveraging LiDAR in a multi-modal framework. If LiDAR is not available in your setup, you may find image-only OCC approaches more suitable.

2. From our limited tests, BEV models trained on surround-view cameras (for both 3D detection and occupancy prediction) show very limited generalization when the number, layout, or intrinsics of the cameras change. The performance tends to drop significantly with such changes.

3. If both data and compute resources are limited, using a LiDAR-only model is generally a better option for transfer.