There were many fine papers on Day 2. In addition to those highlighted by Albert, the invited paper from Vladlen Koltun of Intel on creating good quality images from raw sensor data using neural nets instead of the traditional ISP was quite interesting. He reported the work stemming from several collaborations. These networks could start from raw pixel data and produce denoised images, low light enhanced images, high quality zoomed images, and even sort of decent images from raw DVS data that showed high dynamic range. While it takes a lot of hardware power to produce these images from trained NNs, I think further progress will be made so that someday maybe the ISP will be replaced by NN processing in some cameras, or raw data just sent to the cloud for processing.
And while I am quite biased, the photon-counting imaging papers, i.e. QIS, also showed some very interesting results from several different groups, including getting pretty decent images using just a few (1-3) photons per output pixel on average (Gigajot), motion blur reduction (Univ. Tokyo), and HDR (Purdue).
Together, these papers I am discussing show the growing potential of computational imaging* that may someday disrupt the way photon info is transformed from carriers in a semiconductor (or other material!) into images and video.
*ISP is, of course, another simpler version of computation in imaging.
thanks for sharing info from the workshop. also to Eric, thanks for the longer comment. Will the papers be publically accessible? It seems the papers from past workshops are. How long does it usually take till they are published?
There were many fine papers on Day 2. In addition to those highlighted by Albert, the invited paper from Vladlen Koltun of Intel on creating good quality images from raw sensor data using neural nets instead of the traditional ISP was quite interesting. He reported the work stemming from several collaborations. These networks could start from raw pixel data and produce denoised images, low light enhanced images, high quality zoomed images, and even sort of decent images from raw DVS data that showed high dynamic range. While it takes a lot of hardware power to produce these images from trained NNs, I think further progress will be made so that someday maybe the ISP will be replaced by NN processing in some cameras, or raw data just sent to the cloud for processing.
ReplyDeleteAnd while I am quite biased, the photon-counting imaging papers, i.e. QIS, also showed some very interesting results from several different groups, including getting pretty decent images using just a few (1-3) photons per output pixel on average (Gigajot), motion blur reduction (Univ. Tokyo), and HDR (Purdue).
Together, these papers I am discussing show the growing potential of computational imaging* that may someday disrupt the way photon info is transformed from carriers in a semiconductor (or other material!) into images and video.
*ISP is, of course, another simpler version of computation in imaging.
thanks for sharing info from the workshop. also to Eric, thanks for the longer comment. Will the papers be publically accessible? It seems the papers from past workshops are. How long does it usually take till they are published?
ReplyDeleteYes, they will be published. In the past, they were published 3-4 months after the workshop.
Delete