Could you tell me how to construct a dataset for testing the metrics proposed in a paper?

I am very interested in the use of the ADT dataset mentioned in the paper.
However, I have a confusion: if I have single-object instance files in glb
or other 3D formats, how can I apply them to the SAM3D model? Since the
input format of the SAM3D model is images and their corresponding masks,
which are only in png format. My current idea is to use PyTorch3D to
generate images from various angles (currently selected as 0°, 90°, 180°,
270°) as inputs, then output the predicted glb files, and finally evaluate
the results against the ground-truth glb files. But there is still a
problem: how should the results from these four perspectives be selected?
Should the mean value be taken as the final result? I also doubt whether
the above method is reasonable. Looking forward to your reply.