Improving person re-identification performance by customized dataset and person detection

For person re-identification (re-ID), nearly all person re-ID algorithms use public person re-ID datasets, where these datasets all consist of predefined image crops containing a single person. Unfortunately, these image crops are not optimal for video analysis, so that the person detection becomes suboptimal and person re-ID obtains a lower performance score. In this work, several techniques are presented that customize the person images of a popular public person re-ID dataset.
These techniques consist of customization algorithms based on postprocessing the person-detection bounding boxes using the original frames, resulting in several customized datasets to better facilitate person re-identification. We have evaluated five different ways for customization, based on widening the image crops, various aspect ratios and resolutions, and person instance segmentation. We have obtained a significant increase in performance with widened image crops, yielding a convincing performance increase of nearly 3% in the resulting Rank-1 score. Furthermore, when the applied random-cropping process is further optimized to this customization technique, an increase of even more than 4% is obtained. Both performance gains are a strong indication that any future person re-ID system may benefit from customizations based on the original video frames or from specializing the person detector.