People detection is commonly the first automation step in many computer vision based applications such as 3D body pose estimation and reconstruction. The feature extraction algorithm of Histogram of Oriented Gradient (HOG) combined with supporting vector machine (SVM) has been successful in pedestrian detection and had since inspired a series of variational algorithms that improves its accuracy. However, the extensions and enhancements to HOG on multi-scale is rare. In this paper, we propose HOG detection in non-linearly filtered multi-scale image spaces. Our people detection experiment shows significant detection improvement comparing to the traditional linear multi-scale detections, with recall increase from 26% to 67%, and the precision increase of 3.8 times at the recall of 26%. We conclude that the proposed non-linear multi-scale detection framework is necessary in detecting those objects that are in much larger size than trained detection window.