This repository is a re-implementation of "Real-world Anomaly Detection in Surveillance Videos" with pytorch. As a result of our re-implementation, we achieved a much higher AUC than the original implementation.
Ours full code available at CODE
Download following data link and unzip under your $DATA_ROOT_DIR. /workspace/DATA/UCF-Crime/all_rgbs
- Directory tree
DATA/
UCF-Crime/
../all_rgbs
../~.npy
../all_flows
../~.npy
train_anomaly.txt
train_normal.txt
test_anomaly.txt
test_normal.txt
python main.py
METHOD | DATASET | AUC |
---|---|---|
Original paper(C3D two stream) | UCF-Crimes | 75.41 |
RTFM (I3D RGB) | UCF-Crimes | 84.03 |
Ours Re-implementation (I3D two stream) | UCF-Crimes | 84.45 |
![]() | ![]() |
This code is heavily borrowed from Learning to Adapt to Unseen Abnormal Activities under Weak Supervision and AnomalyDetectionCVPR2018.