Anomaly detection in surveillance videos using deep multiple instance learning (MIL).
This is an implementation of this paper:
Real-world Anomaly Detection in Surveillance Videos
Sultani, Waqas and Chen, Chen and Shah, Mubarak,
The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018
This repo is an updated, refactored version of the original implementation, rewritten in Python3 only (original code was in Matlab and Python2).
- Python 3.6+
- C3D v1.0
- Keras 2.2.4 with Tensorflow 1.9.0 backend (tested on these versions only)
Follow instructions in INSTALL.md to install C3D and its dependencies.
To install Python dependencies run:
pip install -r requirements.txt
To verify that installation was successful run tests:
pytest tests
-
To extract C3D features from test videos and convert them to format accepted by the MIL model adjust
config.ymland run:
python utils/prepare_C3D_features --fast --out_c3d data/c3d/test --out_mil data/mil/test --input_file data/UCF-Anomaly-Detection-Dataset/UCF_Crimes/Anomaly_Detection_splits/Anomaly_Test.txtExtracted raw C3D features will be stored in directory pointed by
--out_c3d. Segmented features, in the format expected by the MIL model will be stored in directory pointed by--out_mil. -
To predict anomaly scores for test videos using the pretrained MIL model (loaded from
pretraineddirectory) run:
python src/predict.py -m pretrained -o scores data/mil/test -
To evaluate predicted scores and compute statistics (ROC, FPR, TPR, etc.) run:
python src/evaluate.py -s scores -o eval_results
-
Firstly extract features from train videos:
python utils/prepare_C3D_features --fast --out_c3d data/c3d/train --out_mil data/mil/train --input_file data/UCF-Anomaly-Detection-Dataset/UCF_Crimes/Anomaly_Detection_splits/Anomaly_Train.txt -
Then, train the MIL model on prepared features and save:
python src/train.py -s pretrained data/mil/train