ASM-Net: Category-level Pose and Shape Estimation Using Parametric Deformation

Created by Shuichi Akizuki (Chukyo University)

Citation

If you find this repository useful in your work, please cite:

@inproceedings{Akizuki_2021_BMVC,
               author = {Akizuki, Shuichi and Hashimoto, Manabu},
               title = {ASM-Net: Category-level Pose and Shape Estimation Using Parametric Deformation},
               booktitle = {British Machine Vision Conference (BMVC)},
               year = {2021}
 }

Introduction

This is a python implementation of ASM-Net: Category-level Pose and Shape Estimation Using Parametric Deformation, a BMVC 2021 paper.

This repository includes:

• Test code of ASM-Net on NOCS REAL275 dataset
• Download link of the pretrained weights
• Active Shape Models for point cloud

Requirements

Our codes have been tested with

• open3d==0.11.1
• opencv-python==4.4.0.44
• torch==1.7.0+cu110
• pytorch3d==0.4.0

Test on NOCS REAL275 dataset

Prepare dataset

Please download the NOCS REAL275 dataset from this repository. We need "Test" of Real dataset, Ground truth pose annotation, and Object meshes. Also, download the detection masks of Mask-RCNN on NOCS REAL275 dataset. Download link

Store these files under asm-net/data/.

Prepare the pretrained weights

Please download pretrained weights and parameters for Active Shape Model (ASM). Download link

Unzip params.zip under asm-net/.

Now, we assume your asm-net repository consist of:

asm-net
├── ASM_Net       ... Network description 
├── NOCS_loader   ... Codes for performance evaluation on NOCS REAL275 dataset.
├── dataset        
|      ├── gts             ... Ground truth pose annotation
|      ├── obj_models      ... Object meshes 
|      ├── real_test       ... "Test" set of NOCS REAL275 dataset
|      └── masks_real_test ... Detection mask by Mask-RCNN.
└── params         
       ├── asm_params  ... Parameters for ASM.
       └── params      ... Pretrained weights of ASM-Net

Run ASM-Net

cd ASM-Net
python test_net_nocs2019_release.py --vis --use_icp

Results are stored under asm-net/ASM_Net/release_proposed_out/real_test_wICP_YYYYMMDDhhmm

Evaluation

cd NOCS_loader
python detect_eval.py --log_dir ../ASM_Net/release_proposed_out/real_test_wICP_YYYYMMDDhhmm/

Reference

Our codes are based on the following repository.

• Open3D-PointNet
• NOCS_CVPR2019