A common assumption in computer vision is that tasks like depth estimation, surface normal prediction, and object recognition are independent problems. A new paper from arXiv challenges that view, arguing that these tasks have meaningful relationships. The authors suggest that knowing one task, such as surface normals, could make another, such as depth estimation, easier.
The paper, titled "Taskonomy: Disentangling Task Transfer Learning," proposes a systematic way to map these relationships. Rather than relying on intuition alone, it aims to build a "taskonomy"—a structured taxonomy of visual tasks that shows which tasks transfer well to others. This could help researchers decide what to train on first or which tasks to learn jointly.
Since this is a single source, there are no conflicting findings to compare. The abstract only sketches the idea, but the core claim is clear: visual tasks are not isolated, and understanding their structure is a valuable research direction.