Top-view Omnidirectional Human Pose Data Synthesis using NeRF

MASTER THESIS - Updating soon

Environment_setup

conda create -n mononerf python=3.8
conda activate mononerf

Model Train

I trained the model on subject number 377 and 387 using the following command. Changing the activation function between Relu, Leaky relu and Elu to find the sufficient results. Also changing the model pipeline to achieve the 3D reconstructed results. Further training and updates will be given soon.

python train.py --cfg configs/human_nerf/zju_mocap/377/adventure.yaml

The model was trained for a total of of 190 epochs, following a total iteration of 97000 approx. Total time for 1 scene training is 24+ hours on a Single Gpu.

Progressive training results

Start of training (1000 iterations)

End of Training (95000 iterations approx)

Rendered output

 python run.py \
    --type movement \
    --cfg configs/human_nerf/zju_mocap/377/adventure.yaml 

out.mp4

More Updates Coming THANK YOU