从单幅图像学习场景深度信息固有的歧义性

The inherent ambiguity in scene depth learning from single images

从单幅图像学习场景深度信息是目前机器学习的一个重要问题,这类方法的原理依据是人类视觉系统可以从单幅图像感知深度.人类单眼感知深度是一种长期进化的能力,其对应的"脑加工机理"仍远未清楚,所以将这种"原理"直接作为"计算依据"的合理性似乎并不充分.事实上,三维场景到二维图像的成像过程,存在严格的几何映射关系.当没有任何先验信息时,场景深度与相机焦距存在固有的歧义性.所以,本文认为目前文献中从单幅图像学习场景深度信息这类方法,由于没有考虑焦距的因素,在原理上存在固有的缺陷.本文通过对实际图像进行测试,证实了这种歧义性的存在.从单幅图像学习深度信息,应该考虑相机内参数的影响,至少是相机有效焦距的影响.

Scene depth inference from a single image is currently an important issue in machine learning, its underlying rationale is the possibility of human depth perception from single images. Human depth perception from single images is the consequence of extended evolution, its underlying brain mechanism is far from well established whereby it seems imprudent to compute the depth based merely on such a fact. In fact, the 3D-to-2D object imaging process must satisfy some strict projective geometric relationship, and without prior knowledge on the camera＇s intrinsics, some ambiguity exists between the scene depth and the camera＇s focal length. We think since the camera＇ intrinsics were not accounted for, the currently reported single-image based depth learning approaches in the literature invariantly contain a crucial deficiency in theory. The ambiguity between the depth and focal-length is also verified by real images. We think in order to increase the accuracy of learning the scene depth from single images, cameras＇ intrinsics should be taken into account, at least the focal-length should also be used as input in both learning and inference phases.