Colour at Scale

There is a lot of conversation around colour forecasting and colour design in architectural forums, trade shows and the like. It is often a hot topic, one which has traditionally been done mostly for marketing purposes and with limited resources and data. Our partner Tarkett invited us to their London showroom during this year’s Clerkenwell Design Week to discuss how we approach this topic at Stylib.

Case study

We decided to explore biases and patterns within colour design, to try and shed light on how free / varied the designs are in terms of colour palettes. While in terms of **composition** and **space quality**, professionals do make a huge difference which is quickly obvious to anyone, is this something we can also say about colour?

Acknowledging a vast market for Colour products and designs, we wanted to address a few questions:

- Is there any sort of “colour bias” in design?

- How “free” are designers and architects when it comes to planning colour?

- How diverse are colour schemes in architecture?

Taking an empirical approach to this question: analyze colour in two contexts: **professional** vs. **amateur:**

- Are colours and palettes very different?

- What colour harmonies are more representative in each group?

- Can we appreciate any distinctive patterns?

We analysed 2 sets of data:

- **Airbnb** listings for the city of Madrid (~25.000 records) - insideairbnb.com

- **Archdaily** interior design projects (~5.000 projects) - archdaily.com

Methodology

A brief note on how we approached this analysis: First, we gathered the data (curated datasets, collected images). Afterwards, we extract the main colours of the image (5 colours and their proportion of presence in the image) by scanning each picture and storing that information. We can then analyze the colour profiles in different ways, for instance comparing histograms for individual components of a given colour space (eg. Red from RGB, more about this below), looking at palettes, colour harmonies, etc.

Comparing individual components

We will use **histograms** to compare different aspects of the colours across the 2 datasets. Histograms are a very common way of understanding colour distribution in photography and the graphic design industry. Here, we will use them to extract some initial insights without having to look at the actual images (remember we’re analysing around 40.000 of them!).

These are normalized, so we can superimpose them, colouring the winner for each section of the histogram. This could yield some interesting insights on individual colour components:

RED

HUE

SATURATION

CHROMA

Color normalization

We can get a basic colour taxonomy by which we can group colours attending to how close they are. This way we could also get counts for colour names and colour palettes.

Colour palettes

Looking at palettes, we can study groups of colours in semantic units, count the most common ones, normalize colours into named ones and study colour harmonies. For our study, a palette is a collection of the most dominant colors in each image (up to 5 colours)

We came up with a visualization to display these palettes adding a notion of proportion in the image (how much “space” they occupy in the image), so we can easily compare palettes at a glance.

Once we compute all palettes we can randomly select some palettes for each dataset and compare them visually to understand potential differences.

Randomly picked palettes from AirBnb:

Randomly picked palettes from Archdaily:

Looking at these two charts, it becomes clear that in both cases, there seems to be an underlying trend in how colour combinations are grouped:

- low saturation (prevalence of greys and desaturated tones)

- prevalence of browns and beiges (as seen in the named colours’ distribution)

- mostly monochromatic or complementary palettes (we discovered that most of the images from both datasets showcase **monochromatic or complementary harmonies, covering around 65%** of the cases)

The differences between datasets are very subtle - they seem interchangeable

Normalized palettes

If we take the previous approach and normalize the colours, we can then obtain palettes of normalized (or named) colours. Since we reduce the number of colors, we can then count how many times each palette appears on each dataset. Here we plot the most common (normalized) palettes from left to right, and the percentage of the dataset that is associated with each one (The proportion of colour is not taken into account for the calculation and it’s shown here to keep a consistent language.)

AIRBNB

For Airbnb, we discovered that **87% of all the images** can be described with these 9 palettes. What is to say that only 6 of the named colours are needed to describe 87% of the images?

49% of all images can be described using just these three colours:

ARCHDAILY

For Archdaily, we discovered that **89% of all the images** can be described with these 9 palettes. What is to say that only 5 of the named colours are needed to describe 89% of the images?

51% of all images can be described using just these three colours:

Discussion

As we have seen, there seems to be very little difference between colour combinations across datasets. This conclusion raises a few interesting questions that deserve further work:

- Are there underlying colour patterns on which interior design constantly relies?

- If the majority of the projects are palette-similar, what are the commercial implications of the ones that stand out?

From image palette to product palette

The previous method was applied to images, without the notion of what appears in the image (so backdrops like skies or objects like plants are taken into account in the calculation). To remove this bias, we took another approach that takes into account the scene categories or objects. This methodology is behind our vertical image search engine, and it allows us to focus on certain categories and discard others, so we can reach more accurate results, in this case taking into account architectural surfaces and objects that can be found in the Stylib’s database.

We can observe how the resulting palettes change significantly :

Cross-referencing this information with what is available in the Stylib database yields some interesting insights. We can take one category (eg. rugs and carpets) and compare uploaded images that contain that category, counting colours, vs. real products from that category inside Stylib. The following plots represent some of these categories in Stylib (where each colour pixel represents a product)

RUGS and CARPETS

TEXTILES (Curtains and upholstery)

Colour moodboards from user images

Thanks to our understanding of the scene, we can extract palettes that are centred around products and not on the scene. We can then plot these palettes by category and create “moodboards” of real-world combinations.

In the next slides, we will plot some palettes that contain a given category. This exercise could help suppliers fine-tune their product stocking strategies according to real market conditions.

The images used for the exercise are taken from Stylib’s user-uploaded images.

MOODBOARDS CONTAINING CARPETS

MOODBOARDS CONTAINING CURTAINS

MOODBOARDS CONTAINING WOODEN FLOORS

Conclusion

In-depth analysis could reveal biases regarding project types, geography, etc.

This data is now available to all of you - who is making the most out of it?

Real-world data analysis yields a competitive edge in understanding the needs of architects and designers and making informed decisions.