The same person and vehicle under different cameras usually undergoes huge appearance variations due to environment illumination, camera settings and viewpoint shifts, which makes it difficult for tasks as tracking, especially in field-of-view non-overlapping camera networks, as shown below.

Following the works employing the widely applied (at that time) camera brightness transfer functions (CBTF), we proposed an bi-directional matching algorithm to pair the instances captured in two cameras with higher accuracy compared with the commonly applied one-way querying.

Moreover, as demonstrated in the figure above, the CBTFs are updated online during the whole process by utilizing the newly paired instances of high confidence. With the self-updating scheme, the accuracy rises higher without supervision.

Also, we collected and prepared both the vehicle and person datasets for our experiments.

The flowchart of the proposed algorithm (in Japanese) is as follows: