Learning feature representations for an object recognition system

摘要

For humanoid robots to be part of our daily lives, not only mobility and their manipulation capability being essential, but their ability to represent and recognize objects in an adaptable manner is also crucial. To this end, we propose an object representation scheme that fits well with the view-based cortical representation of objects found in the primate inferotemporal cortex (IT). We derive our proposal from the simple observation that a single object may exhibit very different sets of visual features when transformed in space. Nonetheless, there are some fixed (object dependent) views of an object, which even with small transformations would not lead to large feature changes. We refer to these views as keyframes. With this in mind, an object is represented with a set of keyframes. The changes in the features around a keyframe are nullified with a neural network that learns to represent the keyframe, and its rotational variations in a compact and rotation invariant form. To evaluate the proposed representation scheme, 100 real life objects are tested in a recognition task. Furthermore, a method for minimizing the number of keyframes for a given object is proposed, which we suggest must yield optimal generalization and computational efficiency. The proposed representation scheme is ideal for building humanoid cognitive architectures as it is decoupled from the recognition system.