Solving Computer Vision Problems Using Traditional and Neural Network Approaches


[MUSIC PLAYING] This is an example of
a synthetic dataset that is publicly available. So you might recognize
the game from the photo. So it’s pretty– not perfect,
but pretty photo-realistic. And the name of that dataset
is “Playing for a Benchmark.” And you can use that
type of synthetic images in order to train your network. And the idea is
that this way, you can get a lot of driving
scenario without having to go outside with a real car. And more importantly, since
you control everything, you know everything in
your synthetic dataset. You know where are
the object, where are your pedestrian, et cetera. So you get the data
notation for free. And this is a very costly path
if you have to do it manually. So having it for free is
really a great benefit. So there is other dataset
that are available. So one can use CARLA, which
is an open-source game engine that you can use to
generate your driving scenario. It’s more basic, but
pretty interesting too. And there is another
publicly available dataset called Synscapes,
which is probably the most photo-realistic,
but a bit restricted in terms of diversity. So the question is,
is it a good idea to use synthetic images in order
to train the network that will be deployed in the real world? So we did a couple
of experiments. And the first question
is, how do we train? [MUSIC PLAYING]

One comment

Leave a Reply

Your email address will not be published. Required fields are marked *