| Summary: | Human beings have a strong capability to identify objects in different viewpoints. Unlike computer vision that requires sufficient training samples in various scales and rotations, biological visual systems can efficiently recognize objects in diverse spatial states. To achieve this objective, images are processed into a segment-based representation and then a vertical stream variational auto-encoder (VSVAE) is utilized to generate images based on the preprocessed segments in this study. The novel structure of the two vertical streams can be also considered as a computational model for the interaction between the ventral pathway and the dorsal pathway in the visual cortex. The reconstructive capability of the VSVAE is testified by using a series of geometric information sets to enhance the segment-based representation. By visualizing the learnt features in the hidden layers of VSVAE, the biological plausibility of the model is discussed. In addition, the proposed methodology is able to facilitate the classification accuracy, especially when the images are severely transformed.
|
|---|
