L4P: Towards Unified Low-Level 4D Vision Perception
This work addresses the need for efficient and general models in computer vision, offering a unified solution for multiple low-level 4D tasks, though it is incremental as it builds on existing pre-trained encoders and task-specific heads.
The paper tackles the problem of specialized architectures for low-level 4D perception tasks by introducing L4P, a unified feedforward model that uses a pre-trained ViT-based encoder with lightweight task heads, achieving competitive performance on tasks like depth estimation and tracking while solving all tasks simultaneously at speeds comparable to single-task methods.
The spatio-temporal relationship between the pixels of a video carries critical information for low-level 4D perception tasks. A single model that reasons about it should be able to solve several such tasks well. Yet, most state-of-the-art methods rely on architectures specialized for the task at hand. We present L4P, a feedforward, general-purpose architecture that solves low-level 4D perception tasks in a unified framework. L4P leverages a pre-trained ViT-based video encoder and combines it with per-task heads that are lightweight and therefore do not require extensive training. Despite its general and feedforward formulation, our method is competitive with existing specialized methods on both dense tasks, such as depth or optical flow estimation, and sparse tasks, such as 2D/3D tracking. Moreover, it solves all tasks at once in a time comparable to that of single-task methods.