Trajectory Recovery: A New Setting and Benchmark for Person Re-identification

Person re-identification can be viewed as a cross-camera retrieval problem, with the assumption that retrieved pedestrians exist in the gallery. However, in actual surveillance scenarios, pedestrians don’t always appear in the camera, and the pedestrians’ temporal and spatial trajectory need be recovered in order to track the pedestrian in the camera network. Motivated by this, we propose a new problem for person re-identification known as “trajectory recovery,” which reconstructs the trajectory of pedestrians in the camera network. Existing datasets lacked the necessary annotated data to address this issue at the same time, so we annotated two new datasets, Market-1501-TR and MSMT17-TR, with each pedestrian’s trajectory annotation. Light variation was also done on both datasets to investigate the effect of light variation on trajectory recovery. Then, we present a multipoint recovery benchmark method that takes into account the relationship between all the query pedestrians and results in an efficient trajectory recovery. To evaluate our method in a new setting, we introduce the degree of trajectory point recovery and the degree of trajectory segment recovery, two new evaluation metrics directly informed by practical surveillance system-level considerations. Experiments on two new datasets demonstrate that our method achieves promising performance and outperforms state-of-the-art methods.