Towards synergistic models of motion information processing in biological and artificial vision. (Vers des modèles synergiques de l'estimation du mouvement en vision biologique et artificielle)

This thesis addresses the study of the motion perception in primates. We propose that scaling up the models rooted in biological vision by taking a task centric approach would gives us further insights to probe biological vision and better constraints to design models. The first part of this thesis relates to a feedforward view of how the motion information is processed in the mammalian brains with specific focus on areas V1 and MT. Based on a standard physiological model describing the activity of motion sensitive neurons in areas V1 and MT, we propose a feedforward model for dense optical flow estimation. This feedforward V1-MT model is benchmarked with modern computer vision datasets and results form a basis to study multiple aspects of dense motion estimation. Benchmarking results demonstrated that a sharp optical flow map cannot be obtained by considering isotropic pooling and motion estimation is disrupted in regions close to object or motion boundaries. It also shows a blindspot in the modelling literature that spatial association of the extracted motion information has not been attempted or has been limited to recovering coarser attributes. In order to improve the motion estimation, we investigated the pooling by MT neurons in terms of spatial-extent and selectivity for integration as well as the decoding strategy in order to obtain a spatially dense optical flow map. We show that by incorporating a pooling strategy that is regulated by form-based cues and considering lateral propagation of the activity, the motion estimation quality is improved. Interestingly, incorporating the form based cues amounts to addition of neurons with different kinds of selectivity to the network. This raises a question, whether or not a minimal network with recurrent interactions in feature domain can exhibit different kinds of feature selectivities or we need to consider explicitly cells with different kinds of selectivity? This question relates to the second part of the thesis. We investigated this question using a ring network model under neural fields formalism with motion direction as feature space, closely mimicing MT physiological experiments. Our model produced a rich variety of results. Our results indicate that a variety of tuning behaviors found in MT area can be reproduced by a minimal network of directionally tuned cells, explicit 2D cues need not be required for motion integration, dynamical changes inthe MT neuronal tuning reported in the literature can be explained through feature domain recurrent interactions and also open the door for accounting transparency by challenging the high inhibition regimes considered by many models in the literature for motion integration. To conclude, we re-emphasize on task-centric modelling approaches and several directions for interfacing studies in biological and computer vision.