Doctoral thesis defense in Computer Science by Ms. Oumaima MOUTIK

The Euromed University of Fez (UEMF) is pleased to inform the public of the

defense of doctoral thesis in “Computer Science”

The thesis defense will take place on Saturday October 12, 2024 at 10:00 a.m. at the UEMF

Location : the Great Hall of the Incubator (LOC001994)

The thesis will be presented by Ms. Oumaima

MOUTIK Under the theme:

“Supervised and self-supervised deep learning techniques for skeleton-based human action recognition in visual media data”

Abstract 

Artificial Intelligence is revolutionizing video understanding applications, seamlessly integrating into our daily routines. However, thus far, action recognition algorithms are incapable of reasoning about videos as humans do. To this end, our thesis aims to develop action recognition solutions that compare with the current state-of-the-art in accuracy but with reduced time and financial costs.

First, we studied the possibility of improving Human-Object Interaction detection by integrating scene information using a compression technique named Knowledge Distillation. This auxiliary task has demonstrated effectiveness in distinguishing between actions. We then considered the surrounding objects, including their motions, to improve Skeleton-based action recognition. This involved designing a set encompassing object information from RGB modality, which was then consistently integrated into the 3-dimensional skeleton data using a novel early fusion technique.

During our thesis journey, we investigated Self-supervised Learning for Skeleton-based action recognition task, driven by the need to mitigate the high cost associated with human annotations. We proposed a novel Pretext Task, "Questions Form Puzzle," based on statistical operations. This work has provided significant results across well-known Skeleton-based datasets and paved the way for extensive future research. Through these contributions, our thesis advances the field of action recognition, striving for human-like video understanding while substantially reducing computational and annotation resource requirements.

Summary 

Artificial intelligence is revolutionizing video understanding applications, integrating perfectly into our daily lives. However, so far, action recognition algorithms are unable to reason about videos the way humans do. To this end, our thesis aims to develop action recognition solutions that compare to the current state of the art in terms of accuracy, but with reduced costs in terms of time and resources.

First, we investigated the possibility of improving the detection of human-object interactions by integrating scene information using a compression technique called knowledge distillation. This auxiliary task demonstrated its effectiveness in distinguishing actions. We then considered surrounding objects, including their movements, to improve skeleton-based action recognition. This involved designing a package integrating object information from the RGB modality, which was then coherently integrated with 3-dimensional skeleton data using a new early fusion technique.

During our thesis journey, we investigated Self-Supervised Learning for skeleton-based action recognition tasks, motivated by the need to reduce the high costs associated with human annotations. We proposed a new pretext task, “Questions Form Puzzle”, based on statistical operations. This work yielded significant results on well-known skeletal-based datasets and paved the way for further in-depth research. Through these contributions, our thesis advances the field of action recognition, striving to achieve human-like understanding of videos, while significantly reducing the requirements for computational and annotation resources.

This thesis will be presented to the jury members:

Share