GDR-Net

This repo provides the PyTorch implementation of the work:

Gu Wang, Fabian Manhardt, Federico Tombari, Xiangyang Ji. GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation. In CVPR 2021. [Paper][ArXiv][Video][bibtex]

News

[2021/08] An extension of this work, SO-Pose by Di et al. (ICCV 2021), has been released (SO-Pose code, mirror).

Overview

Requirements

Ubuntu 16.04/18.04, CUDA 10.1/10.2, python >= 3.6, PyTorch >= 1.6, torchvision
Install detectron2 from source
sh scripts/install_deps.sh
Compile the cpp extension for farthest points sampling (fps):
```
sh core/csrc/compile.sh
```

Datasets

Download the 6D pose datasets (LM, LM-O, YCB-V) from the BOP website and VOC 2012 for background images. Please also download the image_sets and test_bboxes from here (BaiduNetDisk, OneDrive, password: qjfk).

The structure of datasets folder should look like below:

# recommend using soft links (ln -sf)
datasets/
├── BOP_DATASETS
    ├──lm
    ├──lmo
    ├──ycbv
├── lm_imgn  # the OpenGL rendered images for LM, 1k/obj
├── lm_renders_blender  # the Blender rendered images for LM, 10k/obj (pvnet-rendering)
├── VOCdevkit

lm_imgn comes from DeepIM, which can be downloaded here (BaiduNetDisk, OneDrive, password: vr0i).
lm_renders_blender comes from pvnet-rendering, note that we do not need the fused data.

Training GDR-Net

./core/gdrn_modeling/train_gdrn.sh <config_path> <gpu_ids> (other args)

Example:

./core/gdrn_modeling/train_gdrn.sh configs/gdrn/lm/a6_cPnP_lm13.py 0  # multiple gpus: 0,1,2,3
# add --resume if you want to resume from an interrupted experiment.

Our trained GDR-Net models can be found here (BaiduNetDisk, OneDrive, password: kedv).
_{^{(Note that the models for BOP setup in the supplement were trained using a refactored version of this repo (not compatible), they are slightly better than the models provided here.)}}

Evaluation

./core/gdrn_modeling/test_gdrn.sh <config_path> <gpu_ids> <ckpt_path> (other args)