The Hough transform for detecting parameterised shapes in images is still today mostly applied on binary images of contours or connected sets, which implies pre-processing of the images that may be costly and fragile. However the simple estimation of the spatial derivatives provides in every pixel the local geometry that can be used for dense voting processes, directly applied on the gray scale image. For lines and circles, the local information even allows to perform a direct 1-to-1 projection from the image to the parameter space, which greatly accelerates the accumulation process. In this paper we advocate the use of direct detection on gray scale images by combining Hough transform and multi-scale derivatives. We present the algorithms and discuss their results in the case of analytical shapes for order one (lines), and two (circles), and then we present the generalised Hough transform based on quantised derivatives for detecting arbitrary (non-analytical) shapes.