Talks

An overview of deepfake types, what face-swapping techniques there are and why it's dangerous; a review of articles on deepfake detection.

The talk includes an overview of AVS Workgroup and AVS3 video coding standard. The AVS3 key tools, the main features comparison to other state-of-art codecs and various performance test results will be presented, and the deployment of AVS3 and commercial broadcasting in China will be summarized.

AVS3 has been added into DVB/ETSI next generation video codec toolbox standard (DVB-AVC) in July 2022. 

Is it possible to numerically measure video quality? Yes! And VK Calls knows how.

The speaker will touch upon the peculiarities of video transmission during video calls and talk about the team's tools for measuring video quality, including the use of emulation of poor network conditions. He will also present examples of videos and results of their quality measurements using VK Calls methods.

Why and how to write your own player.

During the talk, we will dive into topical issues of neural network optimisation in real-world conditions.

Using vehicle detection on video as an example, the speaker will demonstrate how the right choice of programming language and optimisation methods, including quantisation and the use of hardware-oriented libraries, can significantly increase the performance of neural networks, opening up new opportunities for video analytics.

The speaker will talk about his experience of creating an open source application for real-time camera streaming – go2rtc: why he started developing it, what protocols and problems he faced, what came out of it.

The speaker will present a video compression method with fast frame-by-frame encoding and inter-frame decoding based on compressive sensing. This compression method is about twice as fast as JPEG and x264-intra-ultrafast with PSNR gains of 2.3 and 0.8 dB on average, respectively.

The speaker will talk about the lidar, a laser scanner that sends millions of pulses per second in different directions around itself. By determining the time between the emission of the impulse and its return after reflection, the device, knowing the speed of light, determines the distance to surrounding objects. The result is so-called point clouds, from which high-resolution 3D maps are constructed. This data is used to train neural networks to build optimal trajectories for cars and robots, as well as for direct online obstacle detection. 

The talk will cover the ongoing work in the Media over QUIC (MoQ) IETF working group. What are the purpose and ideas behind one more protocol for media transmission? The speaker will discuss related technologies like QUIC, WebTransport, WebRTC, RTP, HLS, MPEG-DASH and why this is still insufficient.