Skip to content

[CVPR 2024] PyTorch implementation of GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence

License

Notifications You must be signed in to change notification settings

3dlg-hcvc/gigapose

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

33 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

GigaPose: Fast and Robust Novel Object Pose Estimation

via One Correspondence

TL;DR: GigaPose is a "hybrid" template-patch correspondence approach to estimate 6D pose of novel objects in RGB images: GigaPose first uses templates, rendered images of the CAD models, to recover the out-of-plane rotation (2DoF) and then uses patch correspondences to estimate the remaining 4DoF.

The codebase is slightly modified to adapt BOP challenge 2024, if you work on BOP challenge 2023 and have any issues, please go back to previous commits:

git checkout 388e8bddd8a5443e284a7f70ad103d03f3f461c5

News 📣

  • [May 24th, 2024] We added the instructions for running on BOP challenge 2024 datasets and fixed memory requirement issues.
  • [January 19th, 2024] We released the intructions for estimating pose of novel objects from a single reference image on LM-O dataset.
  • [January 11th, 2024] We released the code for both training and testing settings. We are working on the demo for custom objects including detecting novel objects with CNOS and novel object pose estimation from a single reference image by reconstructing objects with Wonder3D. Stay tuned!

Citations

@inproceedings{nguyen2024gigaPose,
    title={GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence},
    author={Nguyen, Van Nguyen and Groueix, Thibault and Salzmann, Mathieu and Lepetit, Vincent},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    year={2024}}

GigaPose's codebase is mainly derived from CNOS and MegaPose:

@inproceedings{nguyen2023cnos,
    title={CNOS: A Strong Baseline for CAD-based Novel Object Segmentation},
    author={Nguyen, Van Nguyen and Groueix, Thibault and Ponimatkin, Georgy and Lepetit, Vincent and Hodan, Tomas},
    booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
    pages={2134--2140},
    year={2023}
}

@inproceedings{labbe2022megapose,
    title     = {MegaPose: 6D Pose Estimation of Novel Objects via Render \& Compare},
    author    = {Labb\'e, Yann and Manuelli, Lucas and Mousavian, Arsalan and Tyree, Stephen and Birchfield, Stan and Tremblay, Jonathan and Carpentier, Justin and Aubry, Mathieu and Fox, Dieter and Sivic, Josef},
    booktitle = {Proceedings of the 6th Conference on Robot Learning (CoRL)},
    year      = {2022},
} 

Installation 👷

Click to expand

Environment

conda env create -f environment.yml
conda activate gigapose
bash src/scripts/install_env.sh

# to install megapose
pip install -e .

# to install bop_toolkit 
pip install git+https://github.com/thodan/bop_toolkit.git

Checkpoints

# download cnos detections for BOP'23 dataset
pip install -U "huggingface_hub[cli]"
python -m src.scripts.download_default_detections

# download gigaPose's checkpoints 
python -m src.scripts.download_gigapose

# download megapose's checkpoints
python -m src.scripts.download_megapose

Datasets

All datasets are defined in BOP format.

For BOP challenge 2024 core datasets (HOPE, HANDAL, HOT-3D), download each dataset with the following command:

pip install -U "huggingface_hub[cli]"
export DATASET_NAME=hope
python -m src.scripts.download_test_bop24 test_dataset_name=$DATASET_NAME

For BOP challenge 2023 core datasets (LMO, TLESS, TUDL, ICBIN, ITODD, HB, and TLESS), download all datasets with the following command:

# download testing images and CAD models
python -m src.scripts.download_test_bop23

For BOP challenge 2024 core datasets (HOPE, HANDAL, HOT-3D), render the templates from the CAD models:

python -m src.scripts.render_bop_templates test_dataset_name=hope

For BOP challenge 2023 core datasets (LMO, TLESS, TUDL, ICBIN, ITODD, HB, and TLESS), we provide the pre-rendered templates (from this link) and also the code to render the templates from the CAD models.

# option 1: download pre-rendered templates 
python -m src.scripts.download_bop_templates

# option 2: render templates from CAD models 
python -m src.scripts.render_bop_templates

Here is the structure of $ROOT_DIR after downloading all the above files (similar to BOP HuggingFace Hub):

├── $ROOT_DIR
    ├── datasets/ 
      ├── default_detections/  
      ├── lmo/ 
      ├── ... 
      ├── templates/
    ├── pretrained/ 
      ├── gigaPose_v1.ckpt 
      ├── megapose-models/

[Optional] We also provide the training code/datasets which is not necessary for testing purposes.

Click to expand
# download training images (> 2TB)
python -m src.scripts.download_train_metaData
python -m src.scripts.download_train_cad 
python -m src.scripts.download_train 

# render templates ( 162 imgs/obj takes ~30mins for gso, ~20hrs for shapenet)
python -m src.scripts.render_gso_templates 
python -m src.scripts.render_shapenet_templates  

If you have training datasets pre-downloaded, you can create a symlink to the folder containing the datasets by running:

ln -s /path/to/datasets/gso $ROOT/datasets/gso

[Optional] Trick for faster converging of ISTNetwork (in-plane, scale, translation): using pretrained weights of LoFTR after Kaiming initialization. Please download the weights and put them in $ROOT_DIR/pretrained/loftr_indoor_ot.ckpt.

Testing on BOP datasets 🚀

If you want to test on BOP challenge 2024 datasets, please follow the instructions below:

Click to expand
  1. Running coarse prediction on a single dataset:
# for 6D detection task
python test.py test_dataset_name=hope run_id=$NAME_RUN test_setting=detection

# for 6D localization task (for only core19 datasets)
python test.py test_dataset_name=lmo run_id=$NAME_RUN test_setting=localization
  1. Running refinement on a single dataset:
# for both 6D detection task
python refine.py test_dataset_name=hope run_id=$NAME_RUN test_setting=detection

# for 6D localization task (for only core19 datasets)
python refine.py test_dataset_name=lmo run_id=$NAME_RUN test_setting=localization

Quantitative results on 6D detection task on HOPEv2 datasets:

Method Refinement Model-based unseen
GigaPose -- 22.57
GigaPose MegaPose --
  1. Evaluating with BOP toolkit:
export INPUT_DIR=DIR_TO_YOUR_PREDICTION_FILE
export FILE_NAME=NAME_PREDICTION_FILE
cd $ROOT_DIR_OF_TOOLKIT
python scripts/eval_bop24_pose.py --results_path $INPUT_DIR --eval_path $INPUT_DIR --result_filenames=$FILE_NAME

If you want to test on BOP challenge 2023 datasets, please follow the instructions below:

Click to expand

GigaPose's coarse prediction for seven core datasets of BOP challenge 2023 is available in this link. Below are the steps to reproduce the results and evaluate with BOP toolkit.

  1. Running coarse prediction on a single dataset:
python test.py test_dataset_name=lmo run_id=$NAME_RUN
  1. Running refinement on a single dataset:
python refine.py test_dataset_name=lmo run_id=$NAME_RUN
  1. Running all steps for all 7 core datasets of BOP challenge:
python -m src.scripts.eval_bop
  1. Evaluating with BOP toolkit:
export INPUT_DIR=DIR_TO_YOUR_PREDICTION_FILE
export FILE_NAME=NAME_PREDICTION_FILE
python bop_toolkit/scripts/eval_bop19_pose.py --renderer_type=vispy --results_path $INPUT_DIR --eval_path $INPUT_DIR --result_filenames=$FILE_NAME

Pose estimation from a single image on LM-O 😺

Click to expand

If you work on this setting, please first go back to previous commits:

git checkout 388e8bddd8a5443e284a7f70ad103d03f3f461c5

Then, download CNOS's detections:

# download gigaPose's checkpoints 
python -m src.scripts.download_cnos_bop23

To relax the need of CAD models, we can reconstruct 3D models from a single image using recent works on diffusion-based 3D reconstruction such as Wonder3D, then apply the same pipeline as GigaPose to estimate object pose. Here are the steps to reproduce the results of novel object pose estimation from a single image on LM-O dataset as shown in our paper:

  • Step 1: Selecting the input reference image for each object. We provide the list of reference images in SuppMat.
  • Step 2: Cropping the input image (and save the cropping matrix for recovering the correct scale for reconstructed 3D models).
  • Step 3: Reconstructing 3D models from the reference images using Wonder3D. Note that the output 3D models are reconstructed in the coordinate frame of input image and in orthographic camera.
  • Step 4: Recovering the scale of reconstructed 3D models using the cropping matrix of Step 2.
  • Step 5: Estimating the object pose using GigaPose's pipeline.

We provide here the inputs and outputs of Wonder3D, here the reconstructed 3D models in Step 1-3 and, this script to recover 3D models in the correct scale. Here is the reconstructed 3D models in the correct scale and in the GT coordinate frame discussed below (note that the GT canonical frame is only for evaluation purposes with BOP Toolkit, while for real applications, we can use the object pose in the input reference image as the canonical frame).

Click to expand

Canonical frame for bop toolkit

For all evaluations, we use bop toolkit which requires the estimated poses defined in the same coordinate frame of GT CAD models. Therefore, there are two options:

  • Option 1: Transforming the GT CAD models to the coordinate frame of the input image and adjust the GT poses accordingly.
  • Option 2: Reconstructing the 3D models, then transforming it to the coordinate frame of GT CAD models by assuming the object pose in the input reference image is known.

Given that the metrics VSD, MSSD, MSPD employed in the bop toolkit depend on the canonical frame of the object, and for a meaningful comparison with MegaPose and GigaPose's results using GT CAD models, we opt Option 2.

Once the reconstructed 3D models are in the correct scale and in the GT coordinate frame, we can now estimate the object pose using GigaPose's pipeline in Step 5:

# download the reconstructed 3D models, test images, and test_targets_bop19.json
mkdir $ROOT_DIR/datasets/lmoWonder3d
wget https://huggingface.co/datasets/nv-nguyen/gigaPose/resolve/main/lmoWonder3d.zip -P $ROOT_DIR/datasets/lmoWonder3d
unzip -j $ROOT_DIR/datasets/lmoWonder3d/lmoWonder3d.zip -d $ROOT_DIR/datasets/lmoWonder3d/models -x "*/._*"

# treat lmoWonder3d as a new dataset by creating a symlink 
ln -s $ROOT_DIR/datasets/lmo/test $ROOT_DIR/datasets/lmoWonder3d/test
ln -s $ROOT_DIR/datasets/lmo/test_targets_bop19.json $ROOT_DIR/datasets/lmoWonder3d/test_targets_bop19.json

# Onboarding by rendering templates from reconstructed 3D models
python -m src.scripts.render_custom_templates custom_dataset_name=lmoWonder3d

# now, it can be tested as a normal dataset as in the previous section
python test.py test_dataset_name=lmoWonder3d run_id=$NAME_RUN
python refine.py test_dataset_name=lmoWonder3d run_id=$NAME_RUN

We provide here the (coarse, refined) results with reconstructed CAD models.

Training

Click to expand
# train on GSO (ID=0), ShapeNet (ID=1), or both (ID=2)
python train.py train_dataset_id=$ID

👩‍⚖️ License

Unless otherwise specified, all code in this repository is made available under MIT license.

🤝 Acknowledgments

This code is heavily borrowed from MegaPose and CNOS.

The authors thank Jonathan Tremblay, Medéric Fourmy, Yann Labbé, Michael Ramamonjisoa and Constantin Aronssohn for their help and valuable feedbacks!

About

[CVPR 2024] PyTorch implementation of GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 99.9%
  • Shell 0.1%