A new part segmentation approach is presented which works on real images. These images contain 3D objects with textures and shadows on a complex background. The proposed approach relies on the outline of the object to guide the grouping of lines using symmetry and colinearity principles. This grouping of lines leads to simple shapes which can be modeled by 3D primitives such as geons or general cylinders. An algorithm implemented on the basis of this approach appears robust to noise and generic conditions. Besides, intermediate-level symmetries employed by the algorithm ensure a good robustness to internal textures and markings. The results obtained demonstrate the validity of the approach as a mean towards 3D generic object recognition from real 2D images.