Relief Impression Image Detection : Unsupervised Extracting Objects Directly From Feature Arrangements of Deep CNN

Traditional sliding windows method for region proposals always focus on the numerical feature values. Instead, we find that the arrangements of convolutional features in deep CNN keep a lot of spatial information of objects in image, and by extracting the positions of high value features as region proposals, we get a unified structure of object detection only based on forward CNN features.Our method has high speed and reasonable performance, since we get the idea from the representations of relievo, so we call this relief impression object detection. We also propose an assumption that the deep detector can act as human’s eye sight tracing without specialized training, we verify this assumption by creating a process called recursive finetuen in our method and get obvious promotion. Our method can get a good perfomace on the limited region proposals situation with very high speed and less computation resource. As far as we know, our work is the first one that notice the massive spatial information stored by the CNN convolutional features arrangement , and use them on the object detection task.