2. ● About me
● What is Action Recognition?
● Data for action recognition
● Pre-DL era models
● DL-era models
● Future prediction
● Q&A
Lecture structure
2
3. ● 10+ years in IT
● 5+ years in AI & CV
● research engineer
● Lviv Polytechnic National University - lecturer, PhD student
● Волонтер в
КомуШоТреба(https://www.facebook.com/komushotreba)
About me
3
5. ● Action classification seeks to assign the correct label (e.g. “cooking,”
“writing,” etc.) to a given image or video.
● Action localization, given a particular action and a video as input, seeks to
identify the correct location and timestamp in the video when the action is
being performed.
5
Action recognition
12. 12
Something-Something v. 2 - 220 000 clips
Sample classes
● Putting something on a surface
● Moving something up
● Covering something with something
● Pushing something from left to right
● Moving something down
27. 27
Tips
● Don’t expect transfer learning to work
● Use robust feature extractors if possible
● The simpler the baseline the better
● Expect to need lots of resources. 1 Nvidia gpu isn’t
enough.
● You’ll need a lot of data. 20 hours won’t be enough.