Abstract
Human lightness perception is remarkably robust against fluctuations in the sensory input – a feature that yields impressive visual illusions and has been subject of extensive psychophysical study for testing perceptual mechanisms proposed to be involved. Modern computational methods additionally have made it possible to investigate how aspects of neural processing might lead to non-veridicality in lightness perception as an emergent property — rather than resulting from specific mechanisms or strategies. Particularly successful is modeling lightness perception using spatial filtering at multiple spatial scales. While not capturing all non-veridicalities, these models are representative of early visual processing and can qualitatively predict perceptual responses to many stimuli. To combine such multiscale spatial filtering models with other proposed mechanisms of lightness perception, we ask which model components have proven useful to explain lightness perception in general (and to what degree), and which components are necessary only when explaining specific phenomena from specific models. We provide a historical roadmap the evolution from earlier to contemporary models, as well as a schematic overview of major differences between models and their effect on resulting model predictions. The schematic overview also made possible an accompanying new, modular, implementation of spatial filtering models. The `multyscale` Python package is fully open-source, with no dependency on proprietary software. It implements several multiscale models, and provides documented examples, demos and tutorials for the models and related topics. The powerful modular implementation is demonstrated, by predicting perceived lightness in a psychophysical paradigm, in detail that would have been difficult and (computationally) time-consuming with previous implementations. Proof-of-concept is given for fitting model parameters to psychophysical data, going beyond the use of previous implementations. Future endeavors can take advantage of the modular nature, by integrating elements of multiscale spatial filtering with either mechanistic approaches (e.g., contour integration) or statistical (e.g., deep learning).