The End-of-End- to-End: A Video Understanding Pentathlon CVPR - PowerPoint PPT Presentation

The End-of-End- to-End: A Video Understanding Pentathlon CVPR workshop, Monday 15 th June 2020

Goal: develop systems that can achieve “higher-level” comprehension of videos (not just objects, but complex actions, events and long-lasting narratives) Computer vision has made tremendous Big Picture progress at core tasks with CNNs trained end- Motivation to-end for classification, detection, segmentation etc. BUT….extending these methods for higher- level tasks is computationally expensive

One way forwards is to use frozen “experts” – representations that have been pretrained on large-scale datasets Enables focus on key questions for higher-level tasks while dramatically reducing compute. Why use - how to make best use of temporal information? pretrained - how to exploit multiple modalities? experts? - how to achieve robustness across different domains? Potential for broader participation outside industry

A video pentathlon challenge – exploring different ways to use pretrained features to solve the Workshop retrieval task objectives A forum for discussing ideas for video understanding

Schedule https://www.robots.ox.ac.uk/~vgg/challenges/video-pentathlon/#schedule

Schedule https://www.robots.ox.ac.uk/~vgg/challenges/video-pentathlon/#schedule

Schedule https://www.robots.ox.ac.uk/~vgg/challenges/video-pentathlon/#schedule

Your mic will be muted (except Workshop when asking questions) Virtual Protocol Use the chat to indicate that you would like to ask a question (we will also try to monitor YouTube)

We are recording this workshop to allow researchers in different timezones to access it later. Recording If you would prefer not to be visible, please turn off your video (and mute) If you want to ask a question, but would prefer not to be in the recording, please let us know afterwards (we will remove this part from the released footage)

Questions for the organisers 1. Use the chat 2. Email: albanie@robots.ox.ac.uk