分割位置提示的可变形部件模型快速目标检测

针对滑动窗口目标检测方法需要穷举搜索目标、检测速度较慢的问题, 提出一种可变形部件模型候选点检测算法.图像先经过两种不同原理的分割方法预处理, 尽量使至少一个分割接近目标真实位置,分割的左上角附近称为候选点. 然后,将可变形部件模型作为底层检测器,模型的训练和测试都只在候选点上进行, 这大大提高了检测速度.在PASCAL 2007数据集上的实验结果表明, 候选点检测在一半类别上的正确率超过了穷举搜索方法.

Fast Object Detection with Deformable Part Models and Segment Locations' Hint

Sliding window detectors need to compute overall scores on all the positions and scales in the image pyramid, which causes the detection speed to be relatively slow. In order to accelerate the detection speed, we propose a candidate points' detection algorithm for deformable part models. Multiple segmentation algorithms are used for each image to generate image segments. The segment's top-left corner is treated as a candidate detection point. We adapt mixture deformable part models as our underlying detectors. The detection operations are only carried on these candidate detection points to accelerate detection speed dramatically. We evaluate the detection performance of our approach on PASCAL 2007 challenge dataset and find that the candidate points' detection is even better than exhaustive search.