Error: principal point is in upper left, not setup for this right now

[rmadan2]

I am running the streetlearn model on stereo images from the KITTI dataset. I have slightly modified demo.py such that it accepts stereo pairs. Furthermore, I have made the following modifications to the intrinsics accepted by the script:
if "matterport" in args.ckpt: intrinsics = np.stack([np.array([[517.97, 517.97, 320, 240], [517.97, 517.97, 320, 240]])]).astype(np.float32) else: intrinsics = np.stack([np.array([[128,128,128,128], [128,128,128,128]])]).astype(np.float32) # intrinsics = np.stack([np.array([[9.786977e+02, 9.717435e+02, 6.900000e+02, 2.497222e+02], [9.786977e+02, 9.717435e+02, 6.900000e+02, 2.497222e+02]])]).astype(np.float32) intrinsics = torch.from_numpy(intrinsics).cuda()
However, The following error message seems to popup:
principal point is in upper left, not setup for this right now
I have tried using both the intrinsics ([128,128,128,128]) and [9.786977e+02, 9.717435e+02, 6.900000e+02, 2.497222e+02]. Is there a way to resolve this?

[crockwell]

Ah, I see. I responded to your other comment in issue #1. I'd recommend moving to DUSt3R. If you want to compare specifically to this model, you may just use Matterport intrinsics. Note the Matterport model is trained only on indoor scenes, so it probably won't generalize well to outdoor scenes.

[rmadan2]

Thank you, I will look further into DUSt3R

[rmadan2]

One thing I do mean to ask, does the network have any sort of memory between the pairs of images fed to it?

[crockwell]

If you're asking how it can process both images jointly, the network concatenates them before passing into the network. So it would scale similarly to passing two "halves" of an image into it. If you're talking about multiple pairs, it has no memory beyond learning from each pair of images via backprop during training.

[rmadan2]

I suppose the reason I ask this is because I only see the error when I give the network a series of stereo pairs (from the KITTI dataset) in a loop. I don't see it when I simply input a stereo pair to demo.py.