City Street

City Street: We collected a multi-view video dataset of a busy city street using 5 synchronized cameras. The videos are about 1 hour long with 2.7k (2704×1520) resolution at 30 fps. We select Cameras 1, 3 and 4 for the experiment (see Fig. 6 bottom). The cameras’ intrinsic and extrinsic parameters are estimated using the calibration algorithm from [52]. 500 multi-view images are uniformly sampled from the videos, and the first 300 are used for training and remaining 200 for testing. The ground-truth 2D and 3D annotations are obtained as follows. The head positions of the first camera-view are annotated manually, and then projected to other views and adjusted manually. Next, for the second camera view, new people (not seen in the first view), are also annotated and then projected to the other views. This process is repeated until all people in the scene are annotated and associated across all camera views. Our dataset has larger crowd numbers (70-150), compared with PETS (20-40) and DukeMTMC (10-30). Our new dataset also contains more crowd scale variations and occlusions due to vehicles and fixed structures.