Impressions from a Computer Vision and Pattern Recognition (CVPR) Newbie
In this article vera.ai's Milica Gerhardt of Fraunhofer IDMT provides first-hand insights into this year's Computer Vision and Pattern Recognition Conference. CVPR is a five-day conference, that was held in Seattle from 17 to 21 June 2024. The conference kicked off with two days of workshops, followed by three days of the main track. Over 12,000 participants including several vera.ai partners attended the event.
As a newbie, I was unprepared for the sheer scale of this event. The first hint came on the plane to Seattle, where many attendees were easily identifiable by their poster carriers. The reality of CVPR's magnitude hit me when I found myself on a lunch break in a massive hall with over 10,000 people. Or when I was swept up in a river of people trying to navigate from the Summit building to Arch, where the poster sessions were held.
The final revelation of CVPR's enormity came with the poster sessions – each lasting 1.5 hours and featuring over 400 posters. Visiting them all meant a brisk 12 seconds per poster. Stopwatch in hand, ready, set, go …
Once the event began, it was clear that CVPR wasn't just massive. It was truly impressive.
The large number of interesting topics and passionate researchers, some staying late after poster sessions to discuss their work, was inspiring. It was an incredible opportunity to gain insights into exceptional research from companies and universities worldwide.
The Expo, where numerous companies set up booths hoping to recruit or connect with future clients, was worth a visit. This was a fantastic opportunity for networking, watching demos, having fun designing fantasy creatures, or getting a portrait done by a robot.
Our highlight was the Media Forensics workshop, an all-day event featuring papers presented in two poster sessions interspersed with four keynote presentations. These talks covered the most recent and significant developments within the media forensics domain, giving us exclusive insights into cutting-edge research.
Justus Thies from Darmstadt University presented state-of-the-art methods for capturing and re-synthesizing software, emphasizing the responsibility that comes with developing algorithms for photorealistic image creation or editing, as these can be misused.
Rebecca Portnoff, Head of Data Science at Thorn, shared her insights on how generative AI models are being manipulated to produce Child Sexual Abuse Material (CSAM), highlighting the potential damage and the steps we can take to build safer AI tools and detect malicious users.
Siwei Lyu from the University of Buffalo showcased an open platform integrating many state-of-the-art algorithms for detecting image, video, and audio deepfakes. He also shared his collected experiences with users of the platform, like those whose accounts were disabled after making suspiciously large numbers of calls, and those who really needed help in their daily fight against fake news.
Lastly, John Collomosse presented a wide variety of work on media provenance with the Adobe Content Authenticity Initiative, which aims to fight fake news and enhance creative control over content reuse in the age of generative AI.
In the poster sessions, vera.ai researchers presented four publications. Our colleagues from the University of Federico II of Naples showcased their latest work on the generalized detection of AI-generated images. The CERTH team presented their paper on image forgery analysis, and Fraunhofer IDMT contributed with one paper on audio provenance analysis and another on audio synthesis detection.
A quick feedback round revealed that all vera.ai participants were positive about their presentations and the feedback they received, marking a successful networking and presentation round for the vera.ai committee.
In conclusion, CVPR was an incredibly inspiring and motivating experience. It is fantastic that media forensics is raising awareness of this vital research area. Thanks to the organization of this workshop, this crucial research field had a significant presence among the giants of CVPR, which is increasingly important in the era of generative AI.
For anyone considering attending this conference, it is highly recommended – don't be discouraged by the high rejection rate. As one oral session presenter humorously thanked the reviewers for accepting their paper after it was previously rejected four times: “Don't give up and see you next year in Nashville!”
Author: Milica Gerhardt (IDMT)
Editor: Anna Schild (DW)
