Neural OmniVideo: Fusing World Knowledge into Smart Video-Specific Models

Projektbeschreibung

Neue Modelle für effektive Videodarstellungen

Bei der Anwendung von Deep Learning auf Bilder mit Rechnersehen wurden große Fortschritte erzielt, doch die Videoanalyse ist aufgrund der Komplexität und Vielfalt von Videodaten schwieriger, da größere Trainingsdatensätze notwendig sind als für Bilder. Video-Rohdaten sind meist hochdimensional, sodass die Verarbeitung gesamter Videos in großem Umfang kostspielig ist. Videospezifische Modelle bieten grundlegende Eigenschaften, allerdings stellen die geringen Informationen in Videos eine Grenze dar, die Funktionen, Anwendbarkeit und Robustheit einschränkt. Im Projekt OmniVideo werden finanziert über den Europäischen Forschungsrat Neural OmniVideo Models entwickelt, indem zwei Ansätze integriert und so die Dynamik von Videos mit Deep-Learning-Rahmen und externen Modellen erfasst werden. Durch die Integration sollen Methodiken zur Videoanalyse und -synthese eingeführt werden, um grundlegend neue und effektive Videodarstellungen zu schaffen.

Ziel

The field of computer vision has made unprecedented progress in applying Deep Learning (DL) to images. Nevertheless, expanding this progress to videos is dramatically lagging behind, due to two key challenges: (i) video data is highly complex and diverse, requiring order of magnitude more training data than images, and (ii) raw video data is extremely high dimensional. These challenges make the processing of entire video pixel-volumes at scale prohibitively expensive and ineffective. Thus, applying DL at scale to video is restricted to short clips or aggressively sub-sampled videos.

On the other side of the spectrum, video-specific models—a single or a few neural networks trained on a single video—exhibit several key properties: (i) facilitate effective video representations (e.g. layers) that make video analysis and editing significantly more tractable, (ii) enable long-range temporal analysis by encoding the video through the network, and (iii) are not restricted to the distribution of training data. Nevertheless, the capabilities, applicability and robustness of such models are hampered by having access to only low-level information in the video

We propose to combine the power of these two approaches by the new concept of Neural OmniVideo Models: DL-based frameworks that effectively represent the dynamics of a given video, coupled with the vast knowledge learned by an ensemble of external models. We are aimed at pioneering novel methodologies for developing such models for video analysis and synthesis tasks. Our approach will have several important outcomes:
• Give rise to fundamentally novel effective video representations.
• Go beyond state-of-the-art in classical video analysis tasks that involve long-range temporal analysis.
• Enhance the perception of our dynamic world through new synthesis capabilities.
• Gain profound understanding of the internal representation learned by state-of-the-art large-scale models, and unveil new priors about our dynamic.