The composition or bed material plays a crucial role in the physical hydromorphological processes of fluvial systems. However, conventional bed material sampling methods provide only pointwise information, which can be inadequate when investigating large rivers of inhomogeneous bed material characteristics. In this study, novel, image-based approaches are implemented to gain areal information of the bed surface composition using two different techniques: monocular and stereo computer vision. Using underwater videos, captured in shorter reaches of the Hungarian Danube River, a comparison of the bed material grain size distributions from conventional physical samplings and the ones reconstructed from the images is carried out. Moreover, an attempt is made to quantify bed surface roughness, using the so-called Structure from Motion image analysis method. Practical aspects of the applicability of image-based bed material mapping are discussed and future improvements towards an automatized mapping methodology are outlined.