Real-World Computer Vision for Real-World Applications: Challenges and Directions

Recent advancements in machine learning, particularly deep learning, have driven the development of next-generation of smart and autonomous applications, especially those that need an in-depth understanding of humans' behaviors and interactions with the physical world. However, there is a gap between current research in computer vision and the inherent real-world limitations of applications. This paper offers holistic solutions embracing real-world computer vision challenges and bringing computer vision to a broad range of applications. The core significance of this paper is creating a holistic privacy-aware ensemble of novel computer vision algorithms and training principles to understand humans' behaviors and interactions with the physical world. To this context, this paper presents multiple fundamental contributions intersecting classical computer vision, deep learning research, and information theory principles. The key contributions include novel privacy-aware identity neural person re-identification, domain-invariant training to bridge the gap between the training data-set and real-world data limitation, enhancing visual resiliency, and knowledge amalgamation across multiple concurrent vision tasks to create full situational awareness.