We are addressing the problem of matching images of scene or of objects when a large collection of reference objects is considered. The paper addresses also the issue of dealing with illumination change and camera position changes. Our approach is firstly based on the use of invariants. Invariants have to be computed locally so that the resulting values will not affected by partial occlusion or accidental highlights. In- variants proved to be a very discriminant piece of information and stored in a hash table they allow efficient indexing of visual shape. Final recognition can be performed using simply a robust voting technique or can be improved using Bayesian decision.