First experience with TF training on 2d correlation (1 batch = 1000 tiles):
On-the-fly Python generation with feed_dict: 1.5 s/batch
Dataset with 1-epoch (400MB) TFRecord: 0.1 s/batch
Epoch in memory: 0.005 s/batch
While we continue to work on the multi-sensor camera (doubling number of sensors for the single-exposure HDR) and the ground truth data for the CNN training, we had fun with capturing aircraft position: blog.elphel.com/2018/05/capt…
Elphel will present new cameras (including the one we are working on now) at #CVPR2018 (booth 132). As this event is held in Salt Lake City we welcome participants to visit Elphel office.
Oleg added driver for the 14MPix MT9F002 sensor and explained how to add other ones. We have 18MPix AR1820 and 12 MPix CMV1200 SFEs in production.
blog.elphel.com/2018/04/impl…
RTL code for the complex lapped transform for the Bayer mosaic color images requires almost 3 times less computational resources than that of the full RGB color data: blog.elphel.com/2018/01/effi…