Historically, the study of visual data has been split into two mirror-image disciplines. Computer vision focuses on analysis—taking raw pixels and extracting semantic meaning, like identifying a car or estimating the depth of a room. Computer graphics focuses on synthesis—taking digital models and rendering them into pixels.
While these fields share the same underlying mathematical foundations of geometry and light transport, they have long operated in isolated silos. As Shi-Min Hu and Piotr Didyk argue in Preface (J. Comput. Sci. Technol. 2025), this fragmentation hinders the development of systems that can perform complex tasks requiring both understanding and generation. The emerging research direction of Computational Visual Media seeks to bridge this gap, creating a unified framework where analysis and synthesis inform one another.
The Duality of Vision and Graphics
The core intuition behind this unification is that vision and graphics are two sides of the same coin. If you can perfectly model how light bounces off a surface (graphics), you are better equipped to "undo" those bounces to understand the surface's true color (vision).
A striking example of this mathematical symmetry was uncovered by Adrián Galdrán and Aitor Alvarez-Gila in On the Duality Between Retinex and Image Dehazing. They demonstrated that Retinex theory—a model for how humans perceive color constancy—is mathematically dual to image dehazing. By recognizing this link, researchers can develop unified algorithms that address multiple types of image degradation simultaneously, rather than treating "fog" and "lighting" as unrelated problems.
Beyond Pixels to Perceptual Understanding
Traditional computer vision often stops at object detection (e.g., "this is a cat"). However, truly intelligent visual media requires a "perceptual-level" understanding that includes aesthetics, quality, and texture.
In UniPercept: Towards Unified Perceptual-Level Image Understanding across Aesthetics, Quality, Structure, and Texture, Yi Xin and Xiangyu Chen propose a framework that moves beyond simple recognition. By training models to reason about how an image feels—its lighting, its composition, its technical quality—they create a "critic" that can guide generative AI to produce more realistic and pleasing results. This represents a shift from machines that just "see" to machines that "perceive."
Bridging Domain-Specific Silos
The fragmentation of visual research isn't just theoretical; it's often trapped within specific application domains. Recent surveys have identified several areas where bridging these silos is critical:
- Game Commentary: Qirui Zheng and Xingbo Wang note in AI-Generated Game Commentary: A Survey and a Datasheet Repository that systems often fail because vision experts (who detect events) and NLP experts (who describe moves) don't collaborate. A unified approach allows an AI to explain why a move matters strategically based on visual cues.
- Persuasive Video: In AI-Empowered Persuasive Video Generation: A Survey, Chang Liu and Han Yu argue that generating high-quality video for e-commerce requires integrating deep learning with classical film theory and psychology.
- Urban Analytics: Shuhan Liu and Yingcai Wu highlight in A survey of urban visual analytics: Advances and future directions that city-scale challenges like traffic congestion require a seamless integration of computational models with interactive human interfaces.
The Path Toward Holistic Visual Systems
The push for unification is driven by the realization that modern visual data is too complex for isolated methods. As Kavita Bala and Kalyan Sunkavalli point out in their editorial on Computational Photography, even the hardware we use is changing. Jointly designing physical optics with computational reconstruction algorithms allows us to "see" through smartphone displays or capture sub-millimeter 3D details.
By breaking down the walls between graphics, vision, and machine learning, the field of Computational Visual Media is moving toward a future where machines don't just process images—they understand the physical and aesthetic laws that govern the visual world. This holistic approach is the key to moving from basic image processing to truly intelligent augmented reality and autonomous systems.