Anomaly-Detection-in-Pose-Estimation

Anomaly Detection in Pose Estimation

In this repo a simple method for anomaly detection in predicted poses is presented

Pose Estimation model

The pose estimation model that is used for extracting the poses is the LightWeight OpenPose

https://github.com/Daniil-Osokin/lightweight-human-pose-estimation.pytorch

Methodology

The methodology is presented in the following steps:

• Pose estimation using the pre-trained model of LightWeight Open Pose (link: https://drive.google.com/file/d/18Ya27IAhILvBHqV_tDp0QjDFvsNNy-hv/view)

• The extracted poses-keypoints are used as input to an anomaly detection Autoencoder model

• According to reconstruction loss the image is characterized as normal or abnormal

Results

Normal	Abnormal	Abnormal

Train

Evaluation

Inference

For testing the pre-trained anomaly detector

Run python autoencoder_results_test_val.py

The inference is performed on some testing images in /test_images directory

Alternatively, it can be run on any directory by modifying the respective line in autoencoder_results_test_val.py https://github.com/mthodoris/PoseAnomalyDetect/blob/ed80173f6b69bcb94f7fc331af346d93274a3033/autoencoder_results_test_val.py#L325