This repo contains information on the Warwick-NTU Multi-camera Forecasting database (WNMF) and baseline multi-camera trajectory forecasting (MCTF) experiments. This repo acompanies the following paper:
Olly Styles, Tanaya Guha, Victor Sanchez, Alex C. Kot, “Multi-Camera Trajectory Forecasting: Pedestrian Trajectory Prediction in a Network of Cameras”, IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2020
Paper link: https://arxiv.org/abs/2005.00282
2020.05.01 - Initial dataset release
2021.08.11 - Cleaned annotations, preprocessing new MCTF problems, more MCTF models, multi-target MCTF preprocessing. Updated annotations are avaiable upon dataset request, and new code is available in the our new repositry, [Trajectory Tensors].
2024.04.13 - Some files in WNMF have become corrupt - these are WNMF_videos/day_20/departures/departure_027.mp4 and WNMF\data\reid_features\departure_features_day_3.npy
If you are interested in downloading the WNMF dataset, please see [this page] to request access.
The data download contains the following:
Videos are paired into entrances and departures. A departure is defined as the 4 seconds before tracking infromation is lost (and the person is therefore assumed to have left the camera view. An entrance is the next camera of re-apperance for this individual. Entrance video clips last for 12 seconds, starting from the moment the individual departed from the other camera view. Each video is processed using [RetinaFace] using an open-source [Pytorch implementation] to mask faces.
Bouding boxes are obtained using an [open-source implementation] of [Mask-RCNN], pre-trained on [MS-COCO]. Individuals are then tracked using an [open-source implementation] of the [DeepSORT] tracking algorithm.
Each track is labelled as as entrance (first frame of the track) or departure (last frame of the track)
RE-ID features are computed using an [open-source implementation] of the [bag-of-tricks] RE-ID model pretrained on [Market-1501].
Cross-camera matches are found using the labelling procedure described in our paper.
Pre-trained weights for each model in our paper.
The camera topology is shown in the figure below.
