The Nature of Similarity in Perceptual Colour Space.

Lead Research Organisation: University of Bristol
Department Name: Experimental Psychology

Abstract

How do we tell how similar two objects are? The nature of similarity is crucial to how we understand the world, and how we generalise to "similar" but not identical situations, but we propose that the nature and geometry of our representations is at least partly determined by what we use them for. If we simply want to discriminate two simultaneously presented coloured patches, then the fundamental limitation is due to photoreceptor noise, and similarity should be (at least partly) determined by how confusable the two patches are in the presence of this noise. Such spaces have been intensively studied, and the powerful techniques of informational geometry have provided insight into the nature of the (low level) representations of space we have. Simple discrimination, though, is not the main use of perceptual signals such as colour, and the nature of the task should also determine the space: if I am simply wanting to discriminate flowers from leaves, then the distance from brown to green should be large (since this is on average discriminatory), whereas that between blue and orange should be short since this discrimination will rarely be important.
The studentship will look at the problem of task dependent representations, how they are learnt, and investigate their characteristics for problems in colour perception. Using the tools provided by information geometry, we will look at both the practical (what colour discriminations are important for given tasks), and the theoretical (how do we characterise the geometry of these spaces). Colour is perfect for this because it is a relatively low dimensional signal, is clearly behaviourally important, but whilst we have a good characterisation of the (receptor based) spaces that characterise simple discrimination, we have a far less good grasp of the representations involved in large colour differences and those used in memory and identification.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/N509619/1 01/10/2016 30/09/2021
2063570 Studentship EP/N509619/1 01/11/2017 07/08/2021 Mubaraka Muchhala
 
Description To date, we have conducted delayed colour estimation tasks on human participants and trained models to classify objects using only object colour.

Our first observation in the delayed colour estimation task is that there are edge effects, where responses to coloured stimuli around the edge of the RGB triangle were biased towards the centre of the triangle. This is because stimuli on the edge of the triangle were restricted in the direction their responses can be biased in.
We therefore fit the data with a 2 component Guassian mixture model, where one component captured some of the bias due to edge effects. This allowed us to observe on a 2D vector graph, the cognitive bias in responses across chromaticity space. There appeared to be basins of attraction, where some areas of the colour space attracted responses (category foci), and some areas of the colour space repelled responses (category boundaries). These findings conform to previous literature, providing evidence for categorical perception across hue and saturation. The categories observed follow those identified by Berlin & Kay as basic colour terms.
A confound in this experiment is that the luminance of the stimuli was not kept constant across the colour space. We are in the process of conducting the experiment again, with constant luminance to ensure that our results are accurate.

Training a model to classify objects using only object colour was a very novel idea. As such, many interesting findings have come out of it. First of all, we have discovered that a model is able to use colour information to identify the probability of an object.
We predicted that the amount of information about object identity provided by colour would equate to log2(11) = 3.46, because Berlin & Kay demonstrated that there are 11 basic colour terms. Therefore 11 may be the optimal number of categories required to communicate about colours in our environment. However, according to our models, colour provided only 0.64 bits of information (or 1.56 colour categories).
We have also observed that the types of categories in the dataset that the model is trained on, has large implications on our data. Objects can vary in how colour diagnostic they are, which influences how much information a colour provides about an object. Objects with highly saturated colours are much less common in our dataset, therefore we found that saturated colours provide more information about objects. Additionally, objects can be categorised at different levels (e.g. furniture, rather than chair, table, bookshelf, etc.), and colour may be more informative for high level categories as opposed to lower levels. In the Imagenet dataset there are many unexpected, low-level categories for which colour may not be a very informative characteristic.
Thus, in further models we plan to use different object categories (Open Images V5 dataset) to observe how this affects the data. We also plan to analyse how the informativeness of colour within the computer model compares with biases in human colour perception from the psychophysics experiment.
Exploitation Route Future research can use this method to observe how the geometry of high-level colour space differs between colour normal and colour vision deficient observers. This idea could also be extended to explore other biases in categorical perception, for example in shape perception or phonetics.
A better understanding of how object classification varies across colour space could also be useful to improve the efficiency of AI colour systems, and to select optimal colour choices for product design.
Sectors Digital/Communication/Information Technologies (including Software),Retail
 
Title Delayed colour estimation task across hue and saturation dimensions. 
Description We have developed a program in Psychopy which displays colour stimuli to participants and collects their estimation of the target colour after a brief delay. Previous researchers have designed a method to test colour estimation, where colour stimuli varied only in one dimension. For example, Bae et al. (2015) displayed colour stimuli in the centre of the monitor screen, and collected responses using a colour wheel surrounding the stimulus which varied continuously in hue. We built upon this design to measure responses to colours varying in 2 dimensions; hue and saturation. We did this by displaying the coloured stimulus on a wheel, and collecting responses using an RGB triangle in CIE1931 space (which varies in hue and saturation), displayed in the centre of the wheel. The participant browsed the RGB triangle with a mouse, which changed the surrounding ring to the corresponding colour. The participant then clicked to submit their response when they were satisfied that the colour of the ring was the same as their memory of the target colour. 
Type Of Material Model of mechanisms or symptoms - human 
Year Produced 2019 
Provided To Others? No  
Impact This novel method enables us to characterise the geometry of high-level colour space. By measuring responses across CIE1931 colour space, we are able to produce a representation of how human colour perception/memory is biased over a perceptually uniform space. In particular, we are able to observe how the pattern of categorical perception, which has been observed for hue, interacts between hue and saturation. This is possible by measuring the difference between the target and response colour for each stimulus, across all participants in xy chromaticity. 
 
Title Biases in human colour perception across hue and saturation. 
Description We collected data using a novel psychophysics experiment. Participants are presented with a coloured ring for 100ms. Colours were sampled from a set of 80 stimuli which varied uniformly across CIE1931 chromaticity space. Following a 900ms delay, the participant uses an RGB triangle to select the colour most similar to their memory of the target colour. We collected the rgb values of the target and response colour for each trial, with 104 measurements per stimulus. 
Type Of Material Database/Collection of data 
Year Produced 2019 
Provided To Others? No  
Impact By converting rgb values into xy chromaticity space, and averaging responses across all participants for each stimulus, we are able to plot a vector graph displaying the difference between the target colour and response colour. Using this data, we can observe how biases in colour perception vary across a 2-dimensional, perceptually uniform colour space. 
 
Title Model trained to classify objects using only object colour 
Description Using the Imagenet dataset, a single rgb value was sampled from each image to represent the colour of that object. We trained a deep neural network to identify object class using only the object colour. This model has a low accuracy rate (1.4%), however it is able to organise objects into their most probable colours, thus performing better than chance. Therefore, providing the model with object colour, provided some meaningful information about the object identity. 
Type Of Material Computer model/algorithm 
Year Produced 2019 
Provided To Others? No  
Impact We are using this model to investigate how much information colour provides about object identity. We use the Jensen-Shannon divergence to measure how changes in colour were reflected by changes in object classification. 
 
Title Replication of Bae, G. Y., Olkkonen, M., Allred, S. R., & Flombaum, J. I. (2015) 
Description We replicated the delayed colour estimation task by Bae et al. (2015). Participants were presented with a colour in the centre of a monitor screen for 100ms. The stimuli were equally spaced in CIELAB space varying in hue. We selected a 130 degree range of stimuli to test (out of the full 360 degree hue circle), which spanned over yellow, orange and green. Following a 900ms delay, participants browsed a continuous colour wheel to estimate the colour closest to their memory of the target colour. We collected the hue angle of the target and response colour. 
Type Of Material Database/Collection of data 
Year Produced 2018 
Provided To Others? No  
Impact Our results replicated the categorical biases observed by Bae et al. (2015). Response variability was not uniform, but rather clustered into a peak at orange/yellow, and a peak at green. As we demonstrated that we could reproduce a categorical effect in this experiment, this program was used as the basis for developing our own novel colour estimation task. We also observed a novel finding in this replication study, where there existed a peak in responses around blue, despite no blue stimuli shown to participants. We discovered that there was a systematic error in responses to green stimuli, with participants responding blue. This was a novel finding, because in previous experiments, the colour range of target stimuli is always equal to the colour range of response stimuli.