Datasets

View on GitHub

Our evaluate human reconstruction on 3 datasets listed below. Please download the datasets from their official sources.

We follow Multi-HMR to prepare 3DPW dataset, execute:

python -m eval.dataset.prepare_3dpw "create_annots()"

Processed 3DPW annotations will be saved to the eval/global_human/annots/ directory.

We follow GVHMR to prepare EMDB dataset.

We follow GVHMR to prepare RICH dataset, please download the following files and place them into the eval/global_human/annots/RICH/ directory:

  • SMPL Annotations: hmr4d_support/rich_test_labels.pt (via GoogleDrive)
  • Camera Ground-Truth: resource/cam2params.pt (via Github)

We also evaluate generic 3D reconstruction (camera Pose and video depth estimation), please follow MonST3R and Spann3R to download, and follow TTT3R to prepare TUM-dynamics and Bonn datasets.

Evaluation

Human Reconstruction

Results will be saved to eval_results/global_human/*.

# You may need to change [--num_processes] to the number of your gpus

# Local human mesh reconstruction - evaluated on 3DPW, EMDB1
CUDA_VISIBLE_DEVICES=0 bash eval/global_human/run.sh

# Global human motion estimation - evaluated on EMDB2
CUDA_VISIBLE_DEVICES=0 bash eval/global_human/run_emdb2.sh

# Global human motion estimation - evaluated on RICH
CUDA_VISIBLE_DEVICES=0 bash eval/global_human/run_rich.sh

Camera Pose Estimation

Results will be saved to eval_results/relpose/*.

CUDA_VISIBLE_DEVICES=0 bash eval/relpose/run.sh # You may need to change [--num_processes] to the number of your gpus and choose sequence length in datasets=('tum_1000')

Video Depth Estimation

Results will be saved to eval_results/video_depth/*.

CUDA_VISIBLE_DEVICES=0 bash eval/video_depth/run.sh # You may need to change [--num_processes] to the number of your gpus and choose sequence length in datasets=('bonn_500')