Status Computer Vision (2) - Mill stained beginning of the heart: a three-dimensional perception
First, it should be noted that, for the most primitive computer for visual representation of data is a digital image - grid (discrete) of the three-dimensional world in a two-dimensional projection plane, if you want to restore the three-dimensional world need to be addressed through a complicated and time-consuming process; the human eye seems to be connected directly to three-dimensional data received. Regardless of how the facts in the end, the two-dimensional digital image is selected to start the computer vision have to make. In addition, based on technology from the digital world to restore the three-dimensional image of multi-view geometry although already mature, but did not receive widespread attention among researchers in computer vision. Currently a variety of segmentation, the algorithm progresses aspects of target detection, tracking, and focus on the two-dimensional digital image processing, thus you can get a glimpse of the current progress in computer vision from a huge distance "come to a complete understanding of the scene," this goal. Need to understand that the primary objective of the current three-dimensional reconstruction based on multi-view geometry of related technologies and did not "come to a complete understanding of the scene" taken into account, only to build a visual model of the real world - grid map model, its production process undergone automatic tie point matching, adjustment to the speed of light, pole (core) to generate a line image, stereo matching dense, fusion point cloud, cloud point network configuration, the texture mapping. The dense stereo matching this step was to generate a discrete point cloud data can represent the three-dimensional world, if you want to get more than one image point cloud data of sight range also need to point cloud integration.
考察我们自己理解场景的两种情形，观察真实的三维世界和观察二维图片，都可以感知到其中的三维信息。这仿佛暗示了基于多视几何的三维感知手段对于“得出场景的完整理解”并不是举足轻重的，其意义或许仅仅在于将真实三维环境与三维环境的透视投影图像区分开来，以防自主行使设备尝试走进一副画里。观察二维图片时，是如何感知三维信息的？基于个人的经验，认为利用了推理这种高级智能。在看到二维图片的一瞬间，人类就可以认出其中包含的目标，目标的二维透视投影形状以及目标在一定照明条件下形成的高光、阴影与三维形状有着对应关系，识别出对应关系就还原了三维信息。透视投影的规律是客观的，平行线的切线消失于灭点是每个人潜意思里的常识，它并没有包含在图片当中。也就是说图片自身对于感知图片的三维信息来说并不是完备的，还需要人类智能利用总结出来的规律和常识做出推理。
透视投影的规律是明确的，应该是计算机视觉的一项基本原理，上面的论述或许缺乏说服力。下面将给出另外一个例子，当观察一幅多山的卫星影像时，第一感觉会将稍暗的一面识别为山的南面，稍亮的一面识别为山的北面，然而却有一条流淌在山顶的河流，村庄和城镇都坐落在山峰的两侧，这是多么的诡异！直觉和认知产生了冲突，为什么会如此呢？因为在人类常处于的环境中，看得见的暗处一定是阴影，阴影处一定是前高后低。然而对于北半球上部朝北的卫星影像来说，是由太阳光从南向北照亮的，所以通常卫星影像上稍暗的一面为山的北面，稍亮的一面为山的南面。经过一定的思维训练，再次观看卫星影像时就可以直接感知到正确的三维地形。这是一个利用外部常识推理来进行三维感知出错的例子，因为所利用的外部常识并不是定律而是经常出现的事实——看得见的暗处一定是阴影，阴影处一定是前高后低。幸好还有其它的常识可以用来发现错误，比如江河绝对不会流淌在山顶上，村落和城镇不会坐落在山峰两侧。
Foregoing default when the human perception of three-dimensional information on two-dimensional picture by first identifying the target, that a fact not known? It has yet to see someone carry out related research. What is the relationship between three-dimensional perception and object recognition is? This is the first loss of computer vision research paper points out that in. Dare guess, three-dimensional perception and object recognition at the same time starts, the initial results have successively, and then began to promote mutual. The humans in the three-dimensional observation of the real world, whether at the same time using the three-dimensional perception perception means and means based on common sense inference based on three-dimensional multi-view geometry of? The relationship between these two means of three-dimensional perception is kind of how? Or we can only guess that the two-dimensional perception means at the same time, and then after target identification based on 3D perception means to perform multi-view geometry, initial results have also experienced successively, and then began to promote mutual. This is a problem of cognitive psychology, with the commencement of this article, will see more deletions associated with cognitive psychology. This would all be a bit ridiculous, after all, each of the participants believe that computer vision computer vision is the inclusion of an integrated discipline of psychology, however, problems related to the psychology of the participants are relying on intuition, not to mention speculation and verification.