Question about camera parameter estimation in 3DTRL

[saltwaterroon]

Hello, thanks for your great work and code release! I've successfully trained 3DTRL on DeiT-S and achieved 79.8% accuracy, which aligns with the results in your paper. However, I've observed an unusual pattern in camera parameter estimation. Rotation and translation tensors appear nearly identical for each input image, as shown in the attached figure. Is this correct? Could you provide insights into this? Have you encountered a similar phenomenon?

[elicassion]

Hi @saltwaterroon ,

I think what you observed is similar to mine. Since we train those estimations without supervision, and the ImageNet dataset is not designed for multi-view inputs, the estimated parameters only have little differences. If you can numerically check them instead of printing, you will notice differences between them.

Again, since it is done without any supervision, we don't expect it to produce exact estimations. Some correlations / differences might be enough for the model.

[saltwaterroon]

Thank you for your assistance. I took your advice and did observe slight numerical differences in the output values, but they were indeed quite minimal. And the rotation tensor tends to approximate zero for most input images. However, in Figure 9 of your visualization, it appears that the camera parameter estimations exhibit significant variation for different input images. Since the differences I've observed are relatively small, I'm curious about the methods or factors that lead to such distinctions in the visual representation. I would appreciate it if you could provide any further clarification.