Moving Pose

Our Presentation (December 2020)

Members

• Andrew Darling
• Eric Hayes
• Mehmet Yilmaz

About

• Given a skeleton based dataset collected from a depth sensor, the goal is to classify certain human actions using the Moving Pose algorithm as well as provide a simple UI.
• To achieve this goal, we implmented the Moving Pose algorithm from the paper stated below and the database stated below.
• This is our Fall 2020 CSCI470 (Introduction to Machine Learning) final Project. CSCI470 is an undergraduate class provided at the Colorado School of Mines. Our team name was: Nestlé.
• Please view /movingpose/gui/README.md to learn more about the GUI and the hardware(s) used.

Weights

The weights used in our implementation of Moving-Pose can be downloaded from HERE. These weights were tuned using cloud GPUs back in December 2020 and are the same weights used in the presentation/demo.

Paper Implemented

• Title: The Moving Pose: An Efficient 3D Kinematics Descriptor for Low-Latency Action Recognition and Detection
• Authors: Mihai Zanfir, Marius Leordeanu, & Cristian Sminchisescu.
• Paper: Zanfir_The_Moving_Pose_2013_ICCV_paper.pdf

Dataset Used

• We used the MSR DailyActivity 3D Dataset dataset: Dataset_Source
• Multiview Action 3D Dataset Action IDs: