If AI is going to make its way out of the chatbox and into our living rooms, it will need to better understand spaces and objects. To further this work, the Allen Institute for Artificial Intelligence Create a huge and diverse database of 3D models of everyday objects, so simulations of AI models can be closer to reality.
Simulators are essentially 3D environments meant to represent real places that a robot or AI might have to navigate or understand. But unlike, say, a modern console game, training simulators are far from realistic and often lack detail, contrast, or interactivity.
Objaverse, as it has a rather awkward but fun name, aims to better that with its collection of over 800,000 (and growing) 3D models with all kinds of metadata. Objects represented range from types of food to tables and chairs to appliances and tools. Anything relatively ordinary you might expect to see in a home, office, or restaurant is represented here.
It is meant to replace legacy object libraries like ShapeNet, which is an outdated database of about 50,000 less detailed models. If the only “light” the AI has ever seen is a generic lamp without a pattern or colour, how do you expect it to recognize a single funky piece of glass or a completely different shape? Objaverse includes differences in common objects so that the model can tell what identifies them despite their differences.
Sure, it probably wouldn’t be necessary for your AI assistant to identify the bookcase as “medieval” or not, but you should definitely know the difference between peeled and unpeeled bananas. But you never know what might matter.
The use of photorealistic images (captured through photogrammetry, obviously) also brings a level of versatility and realism that’s clearly past. Sure, all beds look roughly the same, but what about unmade beds? They are all different!
It’s also useful to have animated objects to do the ‘key stuff’ if you like. Knowing what a fridge, locker, book, laptop, or garage door looks like closed is one thing and another open, but how does it get from A to B? It sounds simple, but if this information is not provided to AI models, it is unlikely to invent or sense it.
You can read more about the features and details of this huge dataset In the AI2 paper you describe. And if you are a researcher, You can start using it now for free via Hugging Face.