A recent trend in model-based object recognition is to build efficient systems for primary hypotheses generation. These systems, also called visual indexing, rely on the assumption that object identification can be performed by recovering local invariants. However, reliable and significant local features are difficult to retrieve when the image background is not uniform.Our solution to this problem takes advantage of a new concept called temporal precedence. The originality of the approach consists in transforming the initial static input image into a dynamic flow of data. Primitives extracted from the image are temporally ranked so as to favor most relevant features for recognition. Most importantly, ranks of features of different kinds may be compared.The architecture of our indexing system is composed of a set of knowledge sources which update in parallel a blackboard structure. The asynchronism produced by the flow of input primitives is used in the activation strategy of knowledge sources.