GraphEnet: Event-driven Human Pose Estimation with a Graph Neural Network

Gaurvi Goyal¹    Pham Cong Thuong¹    Arren Glover¹    Masayoshi Mizuno²    Chiara Bartolozzi¹

¹Istituto Italiano di Tecnologia    ²Sony Interactive Entertainment Inc.   

  

This repository is the official implementation of the paper "GraphEnet: Event-driven Human Pose Estimation with a Graph Neural Network", which was presented at the 2nd Workshop on Neuromorphic Vision: Advantages and Applications of Event Cameras (NeVi2025), International Conference on Computer Vision (ICCV).

To use this setup,

1. copy the data into <<dataDirectory>>/raw.
2. Open main.py and change the data_path variable to <<dataDirectory>>.
3. Run main.py

Installation

To install PyG, follow the instructionfrom here.

For the setup, please install the following libraries:

torch pytorch-lightning torch_geometric matplotlib opencv_python hpe_core albumentations torchvision tensorboard torch-cluster

Dataset

• Please download the eH3.6M and and DHP19 dataset link and organise the downloaded files as follows:

/data
├── datasets
│   ├── eH36M
│   │   ├── EV2       # Contains Ground Truth data
│   │   ├── ledge 
│   │   │   ├── raw   # Contains validation data
│   ├── dhp19
│   │   ├── EV2       # Contains Ground Truth data
│   │   ├── ledge 
│   │   │   ├── raw   # Contains validation data

Training

• To train the GraphEnet model on eh3.6m and dhp19, run the following command:

python3 main.py --node_loss_weight 0.001 --epoch 50 --data_path /data/datasets/dhp19/ledge/ --exp two_weights --dataset dhp19

Evaluation

• To evaluate a single checkpoint, run the following command with --ckpt to specify the checkpoint to be evaluated:

python3 predict.py --data_path /data/datasets/xxxx/ledge/ --ckpt_path path/to/checkpoints/model

Two models are available, one per dataset, in the ckpts folder of the repository.

python3 predict.py --data_path /data/datasets/dhp19/ledge/ --ckpt_path ckpts/single_weight_dhp19.ckpts --arch single_weight --dataset dhp19 --visualise pose