Efficient 3-D Human Pose Estimation: A Synergy of Classical Computer Vision and Deep Learning

This repository is the implementation of a research project for the 2023 Fall Semester Computer Vision class by Team 16, based on the H3WB repository.

Install Dependencies

• We conducted all experiments with Python 3.9 with dependencies listed in requirements.txt.

conda create -n [env name] python=3.9
conda activate [env name]
pip install -r requirements.txt

Data Preparation and Preprocessing

1. Run data.sh in a preferred directory (takes several GB and takes 30~45 minutes.)
2. Put RGBto3D_train.json and RGBto3D_test_img.json to ./data/h3wb/annotations
3. Run python resize.py to resize images to 224x224.
4. Run python split_dataset.py to split the data into pre-defined train, dev, and test sets.

For further details, refer to ./Readme.txt.

Training

We implemented our models in models/ClassicalModel.py and models/CombinedModel.py

output_path=/path/to/model/checkpoint
model_name=resnet50 
# one of {"resnet50", "resnet18"} for baseline models
# one of {"resnet50_4_with_sobel", "resnet18_4_with_sobel"} for sobel operator models

python train.py \
    --learning_rate 1e-5 --batch_size 16 --num_epochs 20 \
    --model_name ${model_name} --use_pretrained \
    --save_path ${output_dir}

Evaluation

checkpoint_path=/path/to/model/checkpoint
python evaluate.py \
    --model_path ${checkpoint_path} --model_name ${model_name}