Download
slide1 n.
Skip this Video
Loading SlideShow in 5 Seconds..
Pose Inference on the Low-Level Features PowerPoint Presentation
Download Presentation
Pose Inference on the Low-Level Features

Pose Inference on the Low-Level Features

186 Vues Download Presentation
Télécharger la présentation

Pose Inference on the Low-Level Features

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -
Presentation Transcript

  1. Pose Inference on the Low-Level Features 上一页表述的对于单个点的匹配分数公式,对于整个局部模板的匹配分数是这样获得的: the computation is very similar to that of chamfer distance.

  2. Representation Using Pose-Adaptive Descriptor 对于人体姿势自适应描述子(Pose-Adaptive Descriptor):它也是基于低层次图像特征的,及图像也是经过提到过的处理方式“加工”过的。 由于边缘轮廓点每幅图都不一样(每个样本数据维度不同),直接利用轮廓点进行分类显然不行。 In order to obtain a unified (constant dimensional) description of images with those different dimensional pose models, and to establish a one-to-one correspondence between contour points of different poses.

  3. Representation Using Pose-Adaptive Descriptor we map the boundary points of any pose model to those of a canonical(典型,就是前面用平行四边形构造的486个整体模型) pose model. 对于图像轮廓取样如下: For human upper bodies (heads and torso), the boundaries are uniformlysampled into eight left side andeight right side locations, and the point correspondence is established between poses based on vertical y coordinatesand side (left or right) information. For lower bodies (legs), boundaries are uniformly sampled into locations vertically with four locations at each y value (inner leg sample points are sampled at 5 pixels apart from outer sample points in the horizontal direction).

  4. Representation Using Pose-Adaptive Descriptor • 对每个采样点取最近的2×2个cell的block作为样本数据(9×4),所以利用PID(pose-insensitive (adaptive) descriptor )所得到的图像特征向量维(8×2+7×4)×36=1584维。

  5. DETECTING AND SEGMENTING MULTIPLE OCCLUDED HUMANS • 论文利用层次局部模板匹配和姿势自适应描述子对局部遮挡问题进行解决。 • 首先利用PID对图像进行粗糙人体目标侦测(通过降低阈值的方式)。our generic detector based on our pose-adaptive features can be used to provide an initial set of human hypotheses (by reducing thresholds to ensure low miss rates)。 • 其次,我们通过层次局部模板树对得到粗略的人体目标假设进行遮挡(occlusion)分析。then more detailed occlusion analysis and optimization can be performed。

  6. DETECTING AND SEGMENTING MULTIPLE OCCLUDED HUMANS • 利用PID和局部模板树确立一些可能的候选人体目标图像块。 • We define a score function Fw (a function of weight vector w) capable of evaluating image responses for any part or part combinations.The function is modeled as weighted sum of individual part responses (matching scores)。

  7. By applying a detection threshold t to the score function Fw, we form seven part or part-combination detectors

  8. In practice, we build a pyramid from the input image and use our sliding-window-based generic human detector to reduce the search space into a small subset and boost it by searching additional hypotheses using the above part/part-combination detectors. We threshold each of the response maps (of full-body matching scores) using a constant global detection threshold t (which is adjustable for trading off precision and recall of detections),merge nearby weak responses to strong responses, and adaptively select modes. 同一张图像的不同尺度的金字塔结构 滑动侦测窗口 图像

  9. This step can also be performed by local maximum selection after smoothing the likelihood image. The union of the maxima forms the set of human hypotheses: • We compute full-body matching scores (assuming no occlusion) for all these hypotheses and denote them by

  10. 根据前面的一部获得可能为人体目标的假设集。我们可以把有遮挡人体目标侦测问题看成一个优化问题,我们可以列出目标函数:根据前面的一部获得可能为人体目标的假设集。我们可以把有遮挡人体目标侦测问题看成一个优化问题,我们可以列出目标函数: 草绿色边框所围成区域即为configuration

  11. We constrain the search space to the initial set of hypotheses, i.e. c belong to u. The goal is to choose the optimal subset of initial hypotheses u and its occlusion ordering so that objective function is maximized. • Given an ordered human configuration c, we can generate its occlusion map Iocc by overlaying regions of global shape estimates。 Iooc

  12. Assuming independence between each observation ui given the configuration c, and treating the template matching scores as pseudolikelihoods, can be decomposed (similar to probability decomposition) as follows:

  13. 对于匹配分数,我们根据Iocc来计算,它所计算的就是假设ui中为被遮挡的部分的匹配分数。对于匹配分数,我们根据Iocc来计算,它所计算的就是假设ui中为被遮挡的部分的匹配分数。

  14. 所以基于以上模型,可以把目标函数写为 • 所以对于任何一个给定排列顺序的人体目标假设都能通过该目标函数计算。

  15. 优化问题 • 对于最大化上一节目标公式的问题是一个NP-难排列组合优化问题。 • 作者给出了一个优化算法:

  16. Combining with Background Subtraction • 结合背景减除,可以很大减少滑动窗口迭代次数,因为剔除背景后的前景图像比原图更小更具体。从而大大提高算法效率。

  17. 实验结果 • 下图表明对于遮挡问题,该算法在某些方面有了较好的解决。

  18. 但是对于有些情况,该算法仍不能很好的检测出来,不如人体运动姿势较大,背景纹理较为丰富的情况下。但是对于有些情况,该算法仍不能很好的检测出来,不如人体运动姿势较大,背景纹理较为丰富的情况下。

  19. Thank you!!!