4DLang · Establishing a Spatio-Temporal Language for Scene Representation

We have recently experienced a boost in AI as the performance of ChatGPT-like large language models has matured from a purely scientific endeavor to deployment in various businesses and real-world applications. Also in computer vision, we have seen tremendous gains that were enabled by scaling to large models trained on vast corpuses of data in an unsupervised fashion. Language is symbolic and can inform about abstract properties and relationships, while vision without human labels does not model explicit semantics and brings distributed representations for spatial structures. Both are complementary, and the fundamental unsolved challenge is to bring them together. The current state of the art is to follow the common paradigm of scale and to naively train models on large amounts of data to exploit the co-occurence of objects in single images and words in text captions to learn their correlation. However, looking at the outputs of these models reveals that they in fact perform extremely poorly in many cases. The next step to approach human-level AI requires reasoning about scenes spatially and semantically at the same time and demands an abstraction of our real world that brings both of these modalities together, while being lightweight and highly efficient. 4DLang presents the solution and introduces a new approach by first creating a primitive-based geometric symbolic abstraction of physical scenes that is then shaped into a spatio-temporal language. It will enable the fine-grained coupling of both modalities and go beyond the state of the art by augmenting large language models with real-world understanding that is only present in observations of moving scenes, as we humans perceive them. This design will fundamentally advance the generalization abilities of AI and have a large impact on downstream applications, such as content interpretation and generation, AI assistants, robotics, and autonomous driving.