Traditional image quality metrics, such as Peak Signal-to-Noise Ratio (PSNR) or Structural Similarity (SSIM), operate under a fundamental assumption: that visual quality can be measured entirely on a flat, two-dimensional plane. These frameworks evaluate pixel-level distortions, blur, and compression artifacts across a single frame.
However, human three-dimensional vision does not work this way. Our perception of stereoscopic media relies on the brain's ability to fuse two slightly offset images—the left and right eye views—into a single, cohesive depth experience.
When stereoscopic images or videos undergo compression or transmission, they don't just suffer from standard 2D artifacts. They can introduce depth distortions, cross-view inconsistencies, and binocular rivalry. An image can remain perfectly sharp and score highly on 2D metrics while still feeling perceptually flat, distorted, or physically uncomfortable to watch. To solve this, researchers have developed stereoscopic quality assessment models that explicitly simulate the neurological and psychological mechanisms of human binocular vision.
Simulating Binocular Fusion and Rivalry
How does the human visual system combine two distinct viewpoints? It does not compute a complex, pixel-by-pixel mathematical disparity map. Instead, the brain processes binocular signals through two primary conceptual pathways: combining matching information and isolating differences.
To capture this computationally without the overhead of heavy disparity estimation, Jiachen Yang et al. developed a full-reference metric in Quality Index for Stereoscopic Images by Separately Evaluating Adding and Subtracting that splits stereoscopic evaluation into two distinct channels:
- Binocular Summation (Adding): This channel adds the signals from both eyes to simulate how the brain reinforces shared visual content, capturing overall luminance and structural clarity.
- Binocular Difference (Subtracting): This channel subtracts one view from the other. By isolating the differences, the model mimics how the brain perceives depth cues and detects binocular rivalry—instances where conflicting information causes visual discomfort.
Left View ───┐
├─── [+] ───> Binocular Summation Channel ───┐
├─── [-] ───> Binocular Difference Channel ──┴─> Predicted Quality
Right View ───┘
In no-reference (or "blind") scenarios, where a pristine original image is unavailable for comparison, modeling this relationship requires evaluating the structural harmony between the two views. The brain expects a high degree of correlation between what the left and right eyes see.
To quantify this without a reference, Huanling Wang et al. introduced a "Left-Right similarity map" in No-reference stereoscopic image-quality metric accounting for left and right similarity map and spatial structure degradation. By measuring how well the structures in the left view map to the right view, their framework can predict whether the brain will smoothly fuse the images or reject them, causing eye strain.
Isolating Depth-Salient Features
Human attention in a 3D environment is highly selective. We do not give equal weight to all parts of a scene; instead, our eyes are naturally drawn to depth transitions—the boundaries where an object separates from its background. If an artifact corrupts a flat, continuous texture, it might go unnoticed. If it corrupts a depth boundary, the entire 3D illusion can collapse.
Recognizing this behavior, Sameeulla Khan Md et al. proposed a method to isolate "depth-salient edges" in Estimating Depth-Salient Edges and Its Application to Stereoscopic Image Quality Assessment. By identifying the specific contours that dictate depth perception and evaluating luminance and depth quality primarily around these regions, their metric mirrors human quality judgments much more accurately than global pixel evaluations.
Standard Edge Detection: Finds all texture changes (leaves, grass, brick patterns).
Depth-Salient Edges: Isolates only the boundaries defining foreground vs. background.
For stereoscopic video, these spatial depth cues must be balanced with temporal comfort. In Objective No-reference Stereo Video Quality Assessment Research, Yang Yu demonstrated that perceived quality is a function of both individual frame fidelity and relative parallax. If the parallax shifts unnaturally between frames, it induces rapid, exhausting muscular adjustments in the viewer's eyes. By combining blockiness distortion metrics with relative parallax measurements, the author showed that no-reference models can successfully predict human visual comfort.
Navigating Immersive and Omnidirectional Spaces
The challenge of modeling stereoscopic quality intensifies when media moves from flat 3D screens to immersive, 360-degree virtual reality (VR) environments. In these spaces, users can look in any direction, meaning models must account for spherical geometry, projection distortions, and localized stitching errors.
Through subjective tracking and eye-movement data compiled in Study of 3D Virtual Reality Picture Quality, Meixu Chen et al. revealed that unique immersive artifacts, like geometric misalignments where viewports stitch together, interact directly with binocular depth perception. Traditional spatiotemporal video quality models—such as those surveyed by Wei Zhou et al. or developed by Zhibo Chen et al.—were built for 2D streams and remain blind to these multi-dimensional spatial conflicts.
To capture how humans perceive quality across multiple viewports simultaneously, Jun Fu et al. designed an Adaptive Hypergraph Convolutional Network (AHGCN) in Adaptive Hypergraph Convolutional Network for No-Reference 360-degree Image Quality Assessment. Instead of treating viewports as isolated images, the hypergraph models complex, multi-directional relationships between different viewing angles, integrating depth-related features to simulate an active human observer scanning a 3D environment.
Finally, to decouple pure depth fidelity from standard 2D compression noise, Zhou Wang et al. introduced the Depth Quality Index (DQI) in Perceptual Depth Quality Assessment of Stereoscopic Omnidirectional Images. DQI isolates the human visual system's sensitivity to color and spatial orientation within omnidirectional depth fields. By separating the evaluation of depth structure from standard luminance artifacts, it ensures that immersive media feels both visually sharp and spatially accurate, providing a true reflection of the user's Quality of Experience (QoE).
Key Foundational Work
Objective No-reference Stereo Video Quality Assessment Research (2013) — contributed 4 concept elements to this direction.
Quality Index for Stereoscopic Images by Separately Evaluating Adding and Subtracting (2015) — contributed 1 concept elements to this direction.