RGB-IR Cross Modality Person Re-Identification
Recognizing persons across images of IR and RGB modalities
Problem Definition
RGB-IR Cross Modality Person Re-Identification (reID) comes under the field of computer vision and the problem has the requirement of matching a person across multiple camera views either from overlapping/non-overlapping cameras of RGB or Infrared (IR) modalities.
Approach
We propose two solutions methods - Attribute Based Representation Learning (ABRL) and Domain Invariant Representation Learning (DIRL) based on novel loss functions (attribute classification loss and domain classification respectively) along with re-ranking for the task of RGB-IR cross modality person reID. We make use of manually annotated common pedestrian attribute information for two large-scale RGB-IR reID datasets as part of ABRL.
Acknowledgements
This project was done as part of my undergraduate’s thesis, you can find the report explaining the methods and experiments here and the repository here. Reach out to me via email for future collaborations in this project.