Fully automatic modeling of image-objects out of example images

研究了图像-物体自动建模的可能性，取得了如下肯定的结果：1.静止图像的目标模型提取假设我们可以通过给一组聚类的颜色区域赋予一个新的名称来很大程度上压缩原始图像，我们可以将该区域集合视为一个对象。与这一原则并列使用几个对象似似性度量，我们可以成功地从彩色图像中全自动地提取卡通、人脸和玩具的结构描述。2.视频中的目标模型提取提出了一种基于背景消除的运动目标提取、边界曲线特征向量提取和基于向量投票的新颖性分析算法。将新的一个存储为新模型，否则仅输出模型标签(=识别结果)。可以提取原始人类动作，即行走和点头等，并用于在线实时模式的进一步识别。3.附加结果1.在分割的图像中找到一组聚类的颜色区域是上述1中最重要的任务之一。我们把这个问题归结为一个图匹配问题，即最大团问题，并提出了两个有效的算法来解决这个问题，这两个算法都利用了图的属性信息。一种是在最大团搜索过程中使用它们，另一种是用于原始图约简。两者都是有效的，但后者主导前者。4.附加结果2.上述2中的系统需要融合的视频流。因此，我们需要对压缩后的视频数据进行解压缩。我们研究了一种直接在压缩数据域进行视频分析的新可能性，并成功地跟踪了MPEG压缩视频中的行走的人。

The possibility of an automatic image-object modeling is studied and affirmative results are attained as follows.1.Object model extraction out of still imagesSuppose we can compress the original image largely by giving a new name to a set of clustered color regions, we may considerthe region set as an object. Using a few object plausibility measures side by side with this principle, we could succeed in extracting, fully automatically, structural descriptions of cartoons, faces, and playthings out of color images. The structural human model could also be extracted out of a still image made up of a few frames of a human walking video.2.Object model extraction out of a videoSe developed an algorithm composed of, the moving object extraction by background elimination, the border curve feature vector extraction, and the novelty analysis by voting from learned vectors to the incoming vector. Novel one is stored as a new model, otherwise only a model label (= recognition result) is output. Primitive human actions, i. e., walking and nodding, etc., could be extracted and used for further recognition in online real-time mode.3.Additional outcome of 1.Finding a set of clustered color regions in a segmented image is one of the most important tasks in 1 above. We reduced this problem into a graph matching problem, i. e., a maximum clique problem and proposed two efficient algorithms to solve it, both exploiting graph attribute information. One uses them in the process of maximum clique search and the other uses them for original graph reduction. Both are effective, but the latter dominates the former.4.Additional outcome of 2.The system in 2 above requires a melted video streams. So we need the decompression of a compressed video data. We examined a new possibility of video analysis directly in a compressed data domain and succeeded to track a walking person in MPEG compressed video.