Abstract : The interaction of shape, material and lighting creates complex patterns of orientation in the image, which play an important role in 3D surface estimation. Depending on the nature of these orientation fields, different cues can be extracted: the orientation fields produced by textured objects carry information about the 3D orientation of the surface (i.e. first order properties), while shading is compressed and stretched depending on the second-order (curvature) properties of the shape. Here, we investigate the relation between shading flows and the perception of shape. We show that by locally modifying the orientations in images of shaded surfaces, we are able to substantially alter the perceived shape. We performed two experiments in which subjects had to match the perceived shape of a test stimulus. In the first experiment, stimuli were generated by adding high frequency patterns (HFPs) on top of diffuse shading. HFPs were oriented along the principal curvature directions of the surface, by 'smearing' noise using line integral convolution. Results show that, depending on the smearing intensity and on the chosen curvature direction (i.e., minimal or maximal), surface features appeared to be either sharpened or rounded respectively. In the second experiment, we used an image processing operator to control the behavior of low-frequencies (LF) shading orientations. By distorting LF gradient patterns on a surface according to curvature information, we are able to precisely control the orientation field in an image. We show that the more a feature is aligned with the shading orientation, the less subjects are able to perceive it correctly. These experiments show that there is a direct relationship between shading orientation flows and our perception of surface curvature, implying that image orientation measurements play a key role in the estimation of shape from shading.