Related Experiment Videos
Invariant object recognition in the visual system with novel views of 3D objects
Simon M Stringer1, Edmund T Rolls
1Oxford University, Centre for Computational Neuroscience, Department of Experimental Psychology, Oxford OX1 3UD, England. simon.stringer@psy.ox.ac.uk
Neural Computation
|November 16, 2002
Summary
This study shows how temporal trace learning in neural networks can achieve 3D object recognition. The model learns object features and their transformations, enabling view-invariant recognition of novel object variations.
Area of Science:
- Neuroscience
- Computer Vision
- Artificial Intelligence
Background:
- Neurons in the inferior temporal cortex may form view-invariant object representations by associating temporally close views.
- Associative learning with short-term temporal memory traces can model this process in neural networks.
- Generalization within views is achieved by neurons learning representations for variations within a single viewpoint.
Purpose of the Study:
- To demonstrate how trace learning can solve in-depth rotation-invariant object recognition.
- To develop representations of geometric feature transforms on 3D objects during rotation.
- To enable recognition of novel 3D object variations from learned feature transformations.
Main Methods:
- Simulations using a hierarchical network model (VisNet) of the visual system.
- Implementing an associative learning rule with a short-term temporal memory trace.
- Developing representations of geometric distortions of surface features under small-angle 3D rotations.
Main Results:
- Trace learning successfully addresses in-depth rotation-invariant object recognition.
- The network learns representations of feature transforms on 3D object surfaces.
- The model demonstrates recognition of novel 3D object variations composed of previously learned features.
Conclusions:
- Trace learning provides a mechanism for developing perspective-invariant representations for 3D object recognition.
- The VisNet model successfully generalizes to novel object configurations after learning feature transformations.
- This approach contributes to understanding how the visual system achieves view-invariant object recognition.