Aerial to Ground One-Shot Vehicle Re-Identification

Source: Deep Learning on Medium

Aerial to Ground One-Shot Vehicle Re-Identification

Vehicle Re-identification is recognizing a vehicle from two different cameras. One-shot vehicle Re-identification is starting the training process with only one example per identity. In this article we went a step further. the cameras used are from different setup. (1) Pedestrian camera: a pedestrian with a smartphone capturing the street. (2) Surveillance camera: a fixed surveillance camera. (3) An aerial source camera.

The radical change in camera position and view port poses multiple challenges. The figure bellow shows the difference between images of the same vehicle from different cameras.

The system we propose is capable of comparing the distribution of features extracted from the input images.

Using a novel anchoring technique we were able to prevent the system under-fitting.