Manifold Learning-Based Feature Transformation for Object Recognition

Recognizing objects is an important task for any computer vision application. Therefore many research works are carried out to make machines understand the objects present in a picture. This paper presents a novel approach to reduce the domain difference of source and target domain using manifold feature learning while preserving the locality and discriminative information of the source data. The method balances the distributional shifts of the domains using MMD metric and the target domain data is maximally varied inorder to avert projecting the data onto trivial dimensions. The manifold feature learning aligns the source and target subspaces by constructing a geodesic flow between the datapoints and integrates the subspaces along the flow. A thorough experimental analysis was performed on the method using benchmark datasets like Office-Caltech 10 using SURF, Decaf6 and NASNETLarge feature descriptors. Various existing methods has not properly exploited the feature aligning techniques for better structuring of the datapoints so that the data can be well positioned in the subspaces. Experimental results shows that the our proposed framework outperforms the traditional domain adaptation approaches on this dataset. Office-Caltech dataset achieved an accuracy of 88.58% for Decaf6 features and 48.21 %for SURF features and 97.24 % for NASNETLarge features respectively.