Music

Hi. I am here as a replacementвор Professor

my and for

sick today and because of that we

we switch the schedule divided so instead

of the best and looked at the eyes

and people talk about object detection

and segmentation today. Maybe a few administrative details at the beginning of the lecture. So to

everyone who has not yet registered for the oral exams, I didn't bring the lists today,

but instead you will have to go to our secretary's office in the blue building to register for your

oral exam and get a time where you can then actually do your oral exam. If you have special

requirements so that none of the dates fit you, there are only very certain cases where we can

make exceptions. For example, if you're an exchange student or something like that that has to leave

the country at a specific time. In these cases, please contact Professor Maia directly and make

an appointment with him directly to clarify that situation. Okay, and then another thing,

the regularization exercise is due this week. To all the students who are in the Wednesday

exercise tomorrow, I will be a little bit short-staffed for the first hour, but in the

second hour I will have some support. So it just might take a little bit longer to do all of the

the assignments or to correct your hand and etc etc. So already sorry about that, but we will do

our best. Okay, so now to today's lecture, object detection and segmentation. You will find that a

lot of the concepts that we will talk today will have appeared one way or another in the lectures

that we had before and that a lot of concepts will be familiar to you. Similarly, for the lectures

that are to come, there again are concepts that you will find represented in the object detection

and segmentation lecture. For example, the encoder decoder structure of a lot of segmentation networks

you will find in autoencoders in the next lecture then that deals with unsupervised learning.

So, this lecture today has two main parts. So we will have a short introduction on what object

detection and segmentation entails. We will then talk specifically about object detection with a

motivation. Why do we want to do that? And then look into different methods that we can use to

detect objects in an image. Specifically, we will look into region-based detectors and single shot

detectors. We will then continue with segmentation, again a short motivation, and then talk about fully

convolutional networks which are very very handy for image segmentation. We will talk about

app sampling which we haven't yet in classification networks and how we can integrate context

knowledge within a segmentation framework. And then look a little bit into more advanced topics

such as both instance segmentation. So object detection and segmentation at the same time.

So to maybe make the terminology a little bit more clear, what we've done so far is we had an image

and we wanted to classify what's in that image. So given that image we wanted we had the label

cat and we wanted our network to predict this is a cat or maybe even cats in a plural sense. Now one

other task that you can give to the network is to recognize the object and localize it. So what is

important here is that we don't have an instance separation. So we have three cats here and all of

them are included in one bounding box. So you can already imagine that this is probably not very

useful. It has been investigated but it's rarely used nowadays and we can look more detailed into

what's contained in the image with actually object detection which will be the first part of this

lecture. So we have here instance recognition and localization as two separate parts of object

detection. So we both want to know what object is in the image and at which position is it and we

want to differentiate between different instances of possibly the same object. You can also imagine

that there would be a dog here in the background that we would want to detect as a separate

instance as well. So we would want to have a bounding box for each instance in the image. Now

then what is very often done is actually semantic segmentation. So we want to segment all parts of

the image that are part of a specific class. So even though this is kind of similar to a semantic