Table 2.

Summary of the image-space human vision simulation methods discussed in this study

Reference	Method used and main innovation	Main limitation
Camp et al. [30]	The authors utilized paraxial ray tracing with corneal topography measurements to compute the on-axis point-spread function (PSF) of the human eye and simulated vision by convolving 2D images with the resulting PSFs	The input ignored the internal aberrations of the eye and the paraxial approach suffered from accuracy issues
Greivenkamp et al. [31]	This work used exact ray tracing with a schematic eye model to calculate the on-axis PSF of the eye and modeled the Stiles-Crawford effect using an apodizing filter	The simulation was limited to 2D images and peripheral vision was ignored
Rokita [32]	The author used repeated filtering with a simple 3 × 3 kernel to approximate the depth-dependent blur of the human eye and utilized the focus distance and input focal length to determine the per-pixel amount of blurring	The lack of real eye information limited the supported types of eye conditions
Barsky [33]	This work used wavefront aberrations to calculate the depth-dependent, on-axis PSFs of the eye and split the input images into depth-dependent slices for convolution with the PSFs	The depth slices had banding artifacts and chromatic aberration and peripheral vision were ignored
Rodríguez Celaya et al. [34]	The authors simulated progressive lenses using sparse, 3D PSF grids (with different axes corresponding to the horizontal angle, vertical angle, and depth), which were interpolated on a per-pixel basis during convolution	The PSF grid was too sparse, the range of incidence angles was limited, and chromatic aberration was ignored
Kakimoto et al. [35]	This algorithm simulated vision through progressive lenses by rendering the scene from multiple views using a precomputed 3D map to compute the per-vertex displacement of each view	The simulation was limited to low-order aberrations and performance scaled poorly with scene complexity
Kakimoto et al. [36]	This work extended the previous multiview method [35] by using conoid tracing to reduce the length of precomputation	This method exhibited the same main limitations as the previous approach [35]
Barsky [37]	The author solved the artifacts of their previous slice-based approach [33] using edge detection to ensure that objects spanning multiple slices are fully included in all slices	Peripheral vision and chromatic aberration were not simulated
Watson [38]	The author used Zernike aberration coefficients and the Fourier transformation to efficiently compute the PSFs of the human eye for varying pupil sizes and object distances	Vision simulation was limited to a single object plane
Tang and Xiao [39]	This work simulated low-order eye aberrations in real-time using an elliptical Gaussian kernel and per-pixel blur field to support peripheral vision and variable eye parameters	Higher-order aberrations (HOA) and chromatic effects were not supported
Barbero and Portilla [40]	The authors used local dioptric matrices to simulate vision through progressive lenses at different gaze directions and approximated the PSFs using samples placed on an ellipse	The inherent eye aberrations and per-pixel depth information were ignored
Cholewiak et al. [41]	This work computed human PSFs by properly simulating longitudinal chromatic aberration	Per-pixel depth information and peripheral vision were ignored
Gonzalez Utrera [42]	The author presented an improved PSF interpolation method for off-axis PSFs and utilized depth-dependent slices to convolve 3D scenes	The PSF grid was too coarse to properly simulate off-axis vision and the slicing caused banding
Csoba and Kunkli [43]	This work used spectacle lens prescriptions to simulate low-order aberrations in real-time environments by utilizing separable complex kernels to approximate the PSFs	HOA were not supported and chromatic aberration and peripheral vision were ignored
Csoba and Kunkli [44]	The authors estimated the physical eye structure from aberrations to compute a coarse PSF grid for varying parameters and simulated vision with an approximately real-time performance profile using tiled convolution and a novel GPU-based per-pixel PSF interpolation approach	The precomputation step was long, peripheral vision was ignored, and partial occlusion was not handled
Lima et al. [45]	This work simulated low-order aberrations using light-gathering trees to efficiently compute refracted light directions for samples on the pupil disk and handled partial occlusion using layered inputs	The simulation was limited to low-order aberrations and peripheral vision was not considered