A 3D product visualisation using Gaussian Splatting. Captured with an iPhone XS in about 5 minutes.
3D in Product Visualisation
Buying products online can be hard. The images don’t properly show what you want to see and if it comes with a 3d model you can already tell that the colours look different than in the photos and the material feels dull and artificial. Which is unfortunate, as without a doubt the best way to observe a 3 dimensional product - if you can’t or don’t want to check it out in real - is to look at a digital 3 dimensional copy. The question arises, what is the reason you seldom encounter a 3d model which looks and feels just like its real counterpart?
3D Definition
Before we answer that question, lets quickly define what we mean when saying „3d“ or „3d model.“
When talking about 3d, we mean a digital 3d model, which is the representation of a real or fictional thing in virtual 3d space. As an observer you have the ability to move around in this space and to look at the model from any angle, at any distance, just as you would in the real world.
Now back to the original question.
Minimise Abstraction
To better understand the problem lets start with some theory.
Anything that is trying to reproduce reality is an abstraction. In case of product visualisations the goal is to minimise the abstraction, so that the customer is not surprised when the product finally arrives at their doorstep and doesn‘t match their expectations - which might be based on a 3d model.
There are two main components that make up a 3d experience, that have to get nailed to create a as real as it gets digital copy:
- the 3d model itself
- the renderer
Different methods go differently about these, but the most common methods will be discussed in the following paragraphs.
Rendering
Rendering techniques differ based on the type of 3d model representation and application. They can be separated into:
Realtime renderer, which are extremely fast, at the cost of realism, and
Offline renderer, which produce much more convincing results, but at the cost of rendering speed. As a result, these renderer are mostly used for movies, architectural visualisations and similar applications that don’t require realtime feedback.
Digital Plastic Arts
Traditionally a 3d artist models the product by hand using images and CAD files, if provided, as reference.
There are plenty of techniques for modelling and a skilled artist will pick the best suited method for the specific circumstances. But all these methods have in common that they start with a state of high abstraction and the more energy the artist puts into them, the smaller gets the abstraction.
You get nothing for free. Similar to plastic arts the modelling process start from nothing (or rather a primitive, common objects are spheres, cubes or similar shapes) and add to it. Each detail has to be added consciously and the more they get refined, the closer it will resemble the real object.
Creating a realistic digital copy of an object that an artist has likely no prior knowledge of, as a result, includes research of shape, functionality, materiality and anything else that is relevant for this specific product. This is usually done by studying provided data, own research and ideally by observing the real product. Only if the object properties are well understood it is possible to create a digital copy with minimal abstraction.
Digital Sculpture
To remain with the analogies to traditional arts: Lets define two viewpoints, from the human perspective and from the renderers perspective. For us humans the real product is in a state of no abstraction, while from the renderers perspective its highly abstract, similar to a block or marble that will become a statue at some point. To minimise the abstraction, the product has to be converted into a format that is desirable for the renderer. While doing so the product gets degraded from the humans perspective. As a result it is best to take as few steps as possible from real to 3d model and to keep the degradation low at every step.
As an example: One possible path is to first translate the product into a digital 2 dimensional space, i.e taking images of it - chiselling the rough shape of our statue into the block of marble - and to produce a digital 3d model out of these in a second step - refining the marble to its final form.
Thats essentially what photogrammetry is doing. Starting from the real thing, taking photos or a video of it and calculating a 3d model from it.
Taking photos is keeping the degradation low when using a high resolution in a well lit environment. The more problematic step is to go from photos to 3d model. This step leads to a pretty high degradation with photogrammetry, in particular with fine details such as fur, thin structures and small perforations. Additionally it does not translate surface properties (reflections, transparency etc.) into the 3d model.
Gaussian Splatting
This is were Gaussian Splatting comes into play. This technique offers an alternative way to convert images into a 3d model. It exceeds in displaying fine detail and captures view dependent properties, such as reflections, as well.
Due to its ability to sustain such a high quality when converting images to a 3d model, it is possible to use a comparably low effort setup while still achieving excellent results. A simple phone video in a well lit environment is able to produce a 3d model with minimal abstraction.
So, what does this mean for e-commerce?
With latest developments, like Gaussian Splatting, it has become possible to create 3d product experiences that closely resemble reality with fairly low effort, making the widespread adoption of photo-real 3d product visualisation much more likely. Ultimately this will make it easier for customers to purchase the product they want - without any guesswork.
Check out our article on how 3D product visualisation is changing online retail.
FAQs
1. What are the challenges in creating 3D models that accurately represent real products?
- Creating digital 3D models that precisely match real products poses challenges such as discrepancies in colour, texture, and material. Despite advancements in technology, current methods sometimes struggle to capture the intricacies of real-world objects.
2. How do different rendering techniques affect the realism of 3D models?
- Realtime renderers prioritise speed but sacrifice realism, while offline renderers produce more convincing results at the expense of rendering speed. Understanding the distinctions between these techniques helps in choosing the appropriate rendering method for specific applications.
3. What is Gaussian Splatting, and how does it improve the creation of 3D models?
- Gaussian Splatting is a technique used to convert images into 3D models with high detail and accuracy, particularly in capturing surface properties such as reflections. Unlike photogrammetry, Gaussian Splatting excels in preserving fine details and minimising degradation, making it a promising approach for creating realistic 3D product visualisations.
