Deep video understanding is a difficult task which requires systems to develop a deep analysis and understanding of the relationships between different entities in video, to use known information to reason about other more hidden information, and to populate a knowledge graph (KG) representation with all acquired information. To work on this task, a system should take into consideration all available modalities (speech, image/video, and in some cases text). The aim of this challenge series is to push the limits of multimodal extraction, fusion, and analysis techniques to address the problem of analyzing long duration videos holistically and extracting useful knowledge to utilize it in solving different types of queries. The target knowledge includes both visual and non-visual elements. As videos and multimedia data are getting more and more popular and usable by users in different domains and contexts, the research, approaches and techniques we aim to be applied in this Grand Challenge will be very relevant in the coming years and near future.

Challenge Overview:

Interested participants are invited to apply their approaches and methods on an extended novel Deep Video Understanding (DVU) dataset being made available by the challenge organizers. The dataset is split into a development data of 14 movies from the 2020-2022 versions of this challenge with a Creative Commons licenses, and a new set of 5 movies licensed from KinoLorberEdu platform. The development data includes: original while videos, segmented scene shots, image examples of main characters and locations, movie-level KG representation of the relationships between main characters, relationships between characters key-locations, scene-level KG representation of each scene in a movie (location type, characters, interactions between them, order of interactions, sentiment of scene, and a short textual summary), and a global shared ontology of locations, relationships (family, social, work), interactions and sentiments. The testing dataset consists of 5 Kinolorber licensed movies.

The organizers will support evaluation and scoring for a hybrid of main query types, at the overall movie level and at the individual scene level distributed with the dataset. Participants will be given the choice to submit results for either the movie-level or scene-level queries, or both. And for each category, queries are grouped for more flexible submission options (please refer to the dataset webpage for more details):

Example Question types at Overall Movie Level:

Multiple choice question answering on part of Knowledge Graph for selected movies.

Fill in the Graph Space - Given a partial graph, systems will be asked to fill in the graph space.

Example Question types at Individual Scene Level:

Find next or previous interaction, given two people, a specific scene, and the interaction between them.

Find a unique scene given a set of interactions and a scene list.

Fill in the Graph Space - Given a partial graph for a scene, systems will be asked to fill in the graph space.

Match between selected scenes and set of scene descriptions written in natural language .

Scene sentiment classification.

A new addition to 2023 challenge is that systems may also submit their results against a secondary dataset where real world noise and various types of perturbations and corruptions are introduced (in visual and audio channels). This will allow the measure of multimodal robustness in this context.


IMPORTANT DATES
DVU development data release: Available now (more updates will be added by April 15)

Testing dataset release : April 15, 2023

Testing queries release: June 2, 2023

Paper submission deadline: July 14, 2023

Submissions of solutions to organizers: July 14, 2023

Results released back to participants: July 24 2023

Notification to authors: July 24, 2023

Camera-ready submission: August 6, 2023

Grand Challenge at ACM Multimedia: TBD