机器视觉问题 - 照片匹配。是解决方案可能/已知，使用OpenCV？ [英] Machine Vision problem - Photo matching. Is a solution possible / known, using OpenCV?

问题描述

在SO上搜索过，并检查了OpenCV列表，但没有找到答案，在这里张贴我的查询。

问题：匹配2张照片场景，从2个稍微不同的摄像机角度拍摄，并且具有稍微不同的镜头失真，略微不同缩放级别，并且在 > 照明条件。

限制：

1. em>稍微不同在上面的语句中可以被认为是最大。在大多数情况下为10％。


2. 所讨论的场景被视为室内场景，或者是有限的室外场景。


3. 匹配精度为75％


4. 照片不高分辨率（用消费级别大多是预算/手机相机拍摄）

什么让我希望这个问题可能是可解决的存在的软件，缝合照片来创建全景。他们似乎自动地找出重叠的部分。即使当地平线方向不完全匹配时，它们也这样做，在曝光水平或背景照明中存在轻微差异，并且存在较小的缩放水平差异。我想，我需要的是一个非常相似的工作流程和一组算法。

请注意，虽然我的问题可能看起来类似一个

解决方案 div>

您需要计算图片之间的单应性其中需要点对应关系，例如 SURF兴趣点。

一旦你有了单应性，你可以做一个图像的投影变换，以使它们匹配。之后，您可以尝试在接缝处进行某种混合，使其看起来无缝。

此纸张描述得很好。您可以将Szeliski使用的多尺度定向补丁替换为SURF兴趣点。以下是一些您可以开始使用的资源：

1. CMU讲座关于单应性和马赛克的


2. 同样的事情，有点羽化


3. 详细PPT

Having searched around on SO, and also checked on OpenCV list but not having found an answer, posting my query here.

Problem: Match 2 photos of the same scene, shot from 2 slightly different camera angles, and with slightly different lens distortions, with slightly different zoom-levels, and shot under slightly different lighting conditions.

Constraints:

1. Slightly different in the above statements can be taken to mean max. of 10% in most cases.
2. The scene in question is to be considered an indoor scene, or an outdoor seen with limited details.
3. Matching accuracy of 75% would be acceptable.
4. The photos aren't high-resolution (shot with consumer grade mostly budget / cell-phone cameras)

What gives me the hope that this problem might be solvable is the existence of software that stitches photos to create panoramas. They seem to figure out the overlapping sections auto-magically. They do so even when the horizon orientations don't match exactly, slight differences exist in exposure level or background illumination, and minor zoom level differences exist. I think, what I need is a very similar workflow and set of algorithms.

Note that while my question might seem similar to one here, actually it is not.

解决方案

You need to compute the homography between the images which needs point correspondences such as SURF interest points.

Once you have the homography, you can do a projective transformation of the images so that they match up. Following that, you can try some sort of blending at the seams to make it look seamless.

This paper describes it pretty well. You can replace the Multi-scale Oriented Patches used by Szeliski with SURF interest points. Here are some more resources to get you started:

1. CMU Lecture on homography and mosaics
2. Same thing with a bit about feathering
3. Detailed PPT

这篇关于机器视觉问题 - 照片匹配。是解决方案可能/已知，使用OpenCV？的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！