Introduction

This code is a pose estimation method based on pixel-wise voting strategy. Our proposed method is improved based on PVNet. So we provide PVNet here for reference(https://github.com/zju3dv/clean-pvnet), and thanks again to Peng et al. for their excellent work.

We made the following change：

• A DDL loss for learning unit vector-field is proposed for PVNet weak constraints

Training and Testing

The code for installation, network training, and testing can be found in the PVNet link in the introduction section. The training and testing commands are shown below:

Take the benchvise as an example

1. Prepare the data related to benchvise:
   

   python run.py --type linemod cls_type benchvise

2. train:
   

   python train_net.py --cfg_file configs/linemod.yaml model yourmodel_dir cls_type benchvise

3. test with the uncertainty-driven PnP on Linemod and OCC-Linemod datasets:
   

   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:./lib/csrc/uncertainty_pnp/lib
   python run.py --type evaluate --cfg_file configs/linemod.yaml model yourmodel_dir cls_type benchvise test.un_pnp True
   python run.py --type evaluate --cfg_file configs/linemod.yaml test.dataset LinemodOccTest model yourmodel_dir cls_type benchvise test.un_pnp True

Schematic