search

RomeroBarata / skeleton_based_anomaly_detection

Code for the CVPR'19 paper "Learning Regularity in Skeleton Trajectories for Anomaly Detection in Videos"
132 stars 49 forks source link

Training on all cameras #24

Closed usamahjundia closed 4 years ago

usamahjundia commented 4 years ago

Im planning to use a custom combination of models for the pose estimation + tracking to create new trajectories and train the model on them because i observed using the aforementioned models together with a customized pipeline on ShanghaiTech dataset shows anomalous examples having smaller error than normal ones.

As for the HR-ShanghaiTech, i could just download the original dataset and filter out the videos using those present in the provided trajectories.

The question is, according to #19 , you used a singular model for all cameras. Question is how did you train it? did you train on 00, then directly used the saved checkpoint and continue training on 01, and so on for all cameras? Or did you use another approach by say, merging all the trajectories assigning frame offsets for different cameras so they dont overlap?

usamahjundia commented 4 years ago

somehow it didnt occur to me that there are 13 folders when theres 12 cameras, for anyone wondering, '00' has all the cameras, haha.