Amazing work! Do you know of any real-world application that is using these or Cohen's Spherical CNNs? I'm thinking of omnidirectional cameras as the source of signal in particular. What would be the training signal if we wanted to do, say, spherical semantic segmentation on a video coming from omnidirectional cams? (where we can't just project the GT labels from a 3D mesh onto a sphere in the case of supervised learning). I guess training on synthetics and somehow bridging the domain gap would be the way to go. But that requires SX data which is usually also hard to get. Also, how does this scale with the resolution?
Amazing work! Do you know of any real-world application that is using these or Cohen's Spherical CNNs? I'm thinking of omnidirectional cameras as the source of signal in particular.
What would be the training signal if we wanted to do, say, spherical semantic segmentation on a video coming from omnidirectional cams? (where we can't just project the GT labels from a 3D mesh onto a sphere in the case of supervised learning). I guess training on synthetics and somehow bridging the domain gap would be the way to go. But that requires SX data which is usually also hard to get.
Also, how does this scale with the resolution?