We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Person re-identification using selective transformation learning.
- Authors
Amin, Fazail; Mondal, Arijit; Mathew, Jimson
- Abstract
Applications like video surveillance, anomaly detection, ego-motion, recognition and re-identification (Re-ID), largely depend upon the ability of the models to learn efficient representations of the input data. Applications like re-identification or similarity matching needs the representations which can handle transformations in the input data in a predictable way. Any change in perspective and viewpoint should not change the identity of the person in re-ID systems and also, it must capture the differences accurately to discriminate two different persons correctly. We propose a Selective Transformation Learning (STL) based model which very efficiently learns to transform the image to obtain the right amount of cropping required to generate feature maps which are invariant to affine transformations of the input image. The STL approach selectively trains each of the spatial transformer modules for specific transformation in an end-to-end framework. Proposed model has very low memory footprint compared to state-of-the-art models yet performs substantially. Compared to the ResNet based or other high capacity models it performs substantially better with such a low capacity. To establish the performance quantitatively it has been tested on three publicly available re-identification datasets and on all the datasets it gives an average of 6% improvement in the mean average precision score as compared to the closest sized state-of-the-art model. This approach can be easily adapted to any other model without any special requirements.
- Publication
Multimedia Tools & Applications, 2023, Vol 82, Issue 25, p38993
- ISSN
1380-7501
- Publication type
Article
- DOI
10.1007/s11042-023-15116-3