Skip to content

Latest commit

 

History

History
50 lines (26 loc) · 1.63 KB

README.md

File metadata and controls

50 lines (26 loc) · 1.63 KB

Introduction

JMLR is an efficient high accuracy face reconstruction approach which achieved Rank-1st of Perspective Projection Based Monocular 3D Face Reconstruction Challenge of ECCV-2022 WCPA Workshop.

Paper in arXiv.

Method Pipeline

jmlr-pipeline

Data preparation

  1. Download the dataset from WCPA organiser and put it at somewhere.

  2. Create cache_align/ dir and put flip_index.npy file under it.

  3. Check configs/s1.py and fix the location to yours.

  4. Use python rec_builder.py to generate cached dataset, which will be used in following steps.

Training

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -u -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=13334 train.py configs/s1.py

Inference Example

python inference_simple.py

Resources

flip_index.npy

pretrained-model

projection_matrix.txt

Results

jmlr-id