POSENET

The algorithm that was put forword in the paper "Simple Baselines for Human Pose Estimation and Tracking or Open Pose", achieves state-of-the-art results on challenging benchmarks. But the problem is with time high time consumption. So is it suitable for Video?

We come up with a modified architecture that can solve the time consumption on the Video processing problems without degrading the performance. For the technical details follow this link to docx-https://lnkd.in/eGSTg6h

We use here the Keypoints generated from our model and use it to compare similarity between 2 videos of poses.

Instructions to use.

To compare a video from previous Keypoint.

Using Python Script

• Use the below command to download the model weight file.

$ gdown --id 17e98AeE1fKUi9_-dwbIxqD3ODTEySP6X 

• Use the below command to run the python program.

$ python <PATH OF THE 'E2E_pose.py' FILE> -v <PATH OF VIDEO TO BE TESTED> -s <PATH OF REFRENCE SOLO FILE> -d <PATH OF REFERENCE DUO FILE> -g <PATH OF REFERENCE GROUP FILE> -w <PATH OF THE MODEL WEIGHT FILE>

Using Notebook

• Upload the files mentioned in the notebook(PoseNet.ipnyb) and follow the steps mentioned in the notebook.

To compare a video from another Video.

Using Notebook

• Upload the files mentioned in the notebook(PoseNet.ipnyb) and follow the steps mentioned in the notebook.

To Generate Keypoint.

Using Notebook

• Upload only the video to generate checkpoints from and follow the steps mentioned in the notebook(PoseNet.ipnyb), and run the "Keypoint Generation" cell.

Results.

Keypoints.

Similarity comparision.

TO DO.

• Improving the similarity comparision algorithm.