Neural Network Implementation for Project Sentinel

Overview

The architecture of the neural network leverages a teacher-student approach for pose estimation, where the teacher model serves as a "ground truth" for our student model.

1. Teacher Network (MMPose)

The teacher network is implemented using MMPose, an open-source toolbox for pose estimation tasks. MMPose provides state-of-the-art models with high accuracy but may require more computational resources.

2. Student Network

TBD

Installation Guide

Setting up MMPose

1. Follow the installation guide on the official MMPose documentation website.

Annotate datasets with MMPose

1. Activate your conda environment with conda activate openmmlab

2. Run the following in your terminal to annotate your file:

   python demo/inferencer_demo.py path/to/video.mp4 --pose2d human --vis-out-dir results_folder --show-progress

Important Notes for MMPose Installation

• Mac Users: Always specify --device cpu in command line calls or device='cpu' in Python code, as CUDA is not available on macOS

• Common Issues:

  • If encountering incompatible mmcv versions, try pip uninstall mmcv followed by mim install "mmcv>=2.0.0,<2.2.0"
  • If encountering incompatible mmdet versions, try pip uninstall mmdet followed by mim install "mmdet>=3.0.0,<3.3.0"
  • Option A takes up more disk space, but tends to be more reliable. If you find yourself struggling to run Option B, try Option A.