Many Perception Tasks are Highly Redundant Functions of their Input Data
This finding challenges assumptions about data representation in perception tasks, potentially simplifying model design, but it is incremental as it builds on existing subspace analysis without introducing new methods.
The paper demonstrates that many perception tasks, such as visual recognition and depth estimation, are highly redundant functions of their input data, as different orthogonal subspaces (e.g., pixel or Fourier domains) can solve these tasks effectively, regardless of whether they capture high or low data variability.
We show that many perception tasks, from visual recognition, semantic segmentation, optical flow, depth estimation to vocalization discrimination, are highly redundant functions of their input data. Images or spectrograms, projected into different subspaces, formed by orthogonal bases in pixel, Fourier or wavelet domains, can be used to solve these tasks remarkably well regardless of whether it is the top subspace where data varies the most, some intermediate subspace with moderate variability--or the bottom subspace where data varies the least. This phenomenon occurs because different subspaces have a large degree of redundant information relevant to the task.