Multi-scale HOG feature used in object detection

Object detection is one of the most popular and difficult field in computer vision. Although deep learning methods have great performance on object detection. For specific application, algorithms which use hand-crafted features are still widely used. One main problem in object detection is the scale problem. Algorithms usually use image pyramid to cover as many scales as possible. But gaps still exist between scale levels in image pyramid. Our work extends some sub scale level to fill the gaps between image pyramids. To this end, we use Gaussian Scales Pyramid to generate sub-scale image and extract HOG feature on the sub-scale. We use framework offered by DPM algorithm and make modification on it. We compare the result of our method with DPM baseline on Pascal VOC database. Our work has great performance on some categories and makes an improvement on the overall performance. This work can be used in other object detection frameworks. We apply multi-scale HOG feature on pre-process procedure of our own detection framework and test it on our own dataset. Then the framework gains performance improvement on precision and recall rate of the pre-process procedure comparing to the original HOG feature architecture.