Vědecké schůzky
čtvrtek 12. 9. 2024 v 10:30
Yulia Gryaditskaya (University of Surrey) – Human-centered approach to generative tasks: Sketch to X (invited talk)
Abstract: A generative AI model that produces high-quality output but lacks user control may be entertaining, but it falls short of practical utility. Text is a widely used input for many modern models targeting image, video, or 3D asset generation. Recently, there has been a growing focus on incorporating spatial inputs, with freehand sketches and doodles emerging as popular modalities. In this talk, I will explore the common challenges of working with sketches and designing effective tools. I will illustrate these challenges through two of our recent projects: (1) a 3D city block massing tool that uses freehand sketches (ECCV’24) and (2) a system for understanding freehand scene sketches (CVPR’24). I will also briefly discuss where we stand in adopting sketches as inputs to generative models.
Speaker info: https://yulia.gryaditskaya.com/
středa 11. 9. 2024 v 10:00
Michal Kučera – Towards Interactive, Robust, and Stereoscopic Style
Transfer (PhD thesis rehearsal)
Abstract: Since its inception in the early 2000s, the research field of style-transfer and automatic stylization has seen a steady rise in popularity up to a point where its algorithms are being employed by professional digital artists in their creation process, allowing them to quickly and conveniently stylize images or video sequences based on either a handmade or a generated example. Even though this research field has seen major strides in recent years, there are still substential issues and limitations preventing larger-scale utilization of such algorithms: limitations such as real-time or interactive stylization of either static images or video sequences, significant quality degradation in cases where example and target keyframes differ too much, temporal coherency of stylized video sequences, infeasible requirements for learning an image-to-image network, as well as stereoscopic applications of style-transfer algorithms remaining uncertain.
In this dissertation thesis we describe the current state-of-the-art in the field of example-based style transfer. Along with that, we propose a set of algorithms that allow interactive production of high quality real-time stylizations of video sequences, both based on semantically meaningful automatic style transfer and keyframe-based learning approaches, on which we introduce new methods to solve the difficult requirement of large paired datasets or domain-specific datasets. We also propose a new method that enables style transfer to still be possible when applied to a stereoscopic scenario.
In particular, we propose: (1) a neural method approximating results of a patch based style transfer method in real time, (2) an interactive method for real-time style transfer of video sequences, (3) a computationally inexpensive method for real-time stylization of facial videos even on low-end devices, (4) a video style-transfer method greatly improving the output quality and long-term coherence, and finally (5) a method able to achieve stereo-consistent style transfer of video sequences.
Combined together, this thesis makes important steps forward to high-quality, realtime, interactive, temporally and stereoscopically consistent style transfer.
středa 17. 7. 2024 v 11:00
Martin Káčerik, Jiří Bittner – SAH-Optimized k-DOP Hierarchies for Ray Tracing
Prezentace článku na HPG 2024
Abstract: We revisit the idea of using hierarchies of k-sided discrete orientation polytopes (k-DOPs) for ray tracing. We propose a method for building a k-DOP based bounding volume hierarchy while optimizing its topology using the surface area heuristic. The key component of our method is a fast and exact algorithm for evaluating the surface area of a 14-DOP combined with the parallel locally ordered clustering algorithm (PLOC). Our k-DOP PLOC builder has about 40% longer build times than AABB PLOC, but for scenes with oblong slanted objects, the resulting BVH provides up to 2.5𝑥 ray tracing speedup over AABB BVH. We also show that k-DOPs can be used in combination with other techniques, such as oriented bounding boxes (OBBs). Transforming k-DOP BVH into OBB BVH is straightforward and provides up to 12% better trace times than the transformation from AABB to OBB BVH.
středa 29. 11. 2023 v 10:00
Vojtěch Radakulan, David Sedláček – The Most Expensive Museum in the World: Three Player Cooperative Game Between VR and PC Platforms Investigating Empathy between Players and Historical Characters.
Prezentace clanku an SIGGRAPH Asia, Art Papers
Abstract: The paper describes an art installation based on a local cooperative cross platform VR/PC game for three players. The story is based on the historical story of a nuclear power plant, which was built, but never turned on because of a public referendum. The game mechanics is built upon interaction of three playable historical characters: engineer, activist, and politician. The main goal was to induce empathy in the players by replaying the game from each character’s perspective. We present the story implementation, custom made interfaces, physical setup, and an evaluation by recordings from four-month display in a public gallery.
středa 4. 10. 2023 v 10:00
Martin Káčerik – PVLI: Potentially Visible Layered Image for Real-Time Ray Tracing
Prezentace clanku z Visual Computer / CGI 2023 (15 min + diskuze)
středa 3. 5. 2023 v 10:00
Ivo Malý – Draw-Cut-Glue: Comparison of Paper House Model Creation in Virtual Reality and on Paper in Museum Education Programme – Pilot Study
Prezentace clanku na EG 2023 – Education (20 min + diskuze)
Ivo Malý, Iva Vachková, David Sedláček. Draw-Cut-Glue: Comparison of Paper House Model Creation in Virtual Reality and on Paper in Museum Education Programme – Pilot Study
Martin Káčerik – On the Importance of Scene Structure for Hardware-Accelerated Ray Tracing
Prezentace clanku na WSCG 2023 – communication papers (10 min + diskuze)
Martin Káčerik, Jiri Bittner – On the Importance of Scene Structure for Hardware-Accelerated Ray Tracing
Asbtract: Ray tracing is typically accelerated by organizing the scene geometry into an acceleration data structure. Hardware-accelerated ray tracing, available through modern graphics APIs, exposes an interface to the acceleration structure builder that builds it, given the input scene geometry. This builder is, however, opaque, with limited control over the internal building algorithm. Additional control is available through the layout of builder input data, a geometry of the scene structured in a user-defined way. In this work, we evaluate the impact of a different scene structuring on the performance of the ray-triangle intersections in the context of hardware-accelerated ray tracing. We discuss the possible causes of significantly different outcomes for the same scene and outline a solution in the form of automatic input structure optimization.