Created by Shuichi Akizuki (Chukyo University)
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}
}
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
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
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/
.
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
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
cd NOCS_loader
python detect_eval.py --log_dir ../ASM_Net/release_proposed_out/real_test_wICP_YYYYMMDDhhmm/
Our codes are based on the following repository.