Is computational imaging the next frontier in computer vision? As materials, lighting, and geometry become more complex an ordinary camera starts to be the limiting step in the vision pipeline. A case in point are transparent objects which cannot be seen by an ordinary camera (which mimics the human eye). However, a plenoptic camera that captures polarization spawns a unique texture to previously invisible objects. The polarized texture results from complex light-matter interactions in shape and refractive index. Such texture can be subsequently modeled and “baked” into neural pipelines to enable inference on transparent objects, a decades-open task in traditional computer vision. WIth polarization simply scratching the surface of what computational imaging can do, we end with a discussion of how broader forms of plenoptic data can be leveraged in future neural pipelines.
Bio: Agastya Kalra is currently a Principal Engineer at Akasha Imaging, where he is also a founding team member. His research interests lie at the intersection of computational imaging, multi-view geometry, and machine learning. He has published in NeuriPS, ICCV, and CVPR and is a co-inventor on more than a dozen patents in computational imaging. His website can be found at kalraa.github.io