CVPR Daily - 2018 - Thursday

properties of whatever it is imaging. It is specially designed for certain tasks in which colour sensing is optimised for that task. He adds that there are already commercial conventional hyperspectral cameras, but they are typically very slow and may suffer from low-spatial resolution, but their camera works like an RGB camera but with a different colour channel. Lin tells us one of the challenges has been how to design the optimal spectral response: “ Now we try to use deep learning to speed up these things. We find that the CCD, the spectral filter part, is actually a special part of the 1 x 1 convolution layer. A basic keystone of the deep learning. We just design our deep learning algorithm and other things directly seeing the whole spectrum of the image. In this way, we can optimise the optimal spectral response function, which is through the learned filter convolution layer. Unlike other cameras, autonomous cars or surgery tools, they just see the RGB. We are actually seeing the whole spectrum of the image .” Lin thinks their camera will be very quickly adopted by industry, because it’s simple and helpful. He points to a heat map showing spectral reconstruction and says that other methods introduce a high amount of error. Red means that error is high, but their map is totally blue. Other people acquire images and then process them, but they work on the image capturing part, increasing performance by 10 or 20 per cent , which is a huge improvement. Summing it up, Antony concludes: “ This camera can see more than what humans see . ” Thursday 21 This camera can see more than what humans see...” Lin Gu Antony Lam