search

WangLongZJU / DeepAC

[ICCV 2023] Deep Active Contours for Real-time 6-DoF Object Tracking
62 stars 6 forks source link

readme

Deep Active Contours for Real-time 6-DoF Object Tracking

Project Page | Paper


Deep Active Contours for Real-time 6-DoF Object Tracking
Long Wang*, [Shen Yan]()*, [Jianan Zhen](), [Yu Liu](), [Maojun Zhang](), Guofeng Zhang, Xiaowei Zhou
ICCV 2023

Installation

conda create -n deepac python=3.8
conda activate deepac
# install torch
pip install torch==1.8.1+cu101 torchvision==0.9.1+cu101 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
# install pytorch3d
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install -c bottler nvidiacub
pip install "git+https://github.com/facebookresearch/pytorch3d.git@v0.7.1"
# install other dependencies
pip install -r requirements.txt

Download the pretrained models, and place them in workspace/train_bop_deepac, like:

|--- workspace/train_bop_deepac
|       |--- logs-2024-01-08-15-52-47
|       |         |--- train_cfg.yml
|       |         |--- model_last.ckpt

Demo

After the installation, you can refer to this page to run the demo with your custom data.

Training and Evaluation

Dataset setup

  1. DeepAC uses eight datasets from the BOP challenge, namely HOPE, IC-BIN, IC-MI, T-LESS, TUD-L, LM, YCB-V, and RU-APC for training and validation. You can download the datasets from here. For evaluation, DeepAC uses RBOT, BCOT and OPT datasets.

  2. DeepAC uses the data SUN2012 to change the background for augmentation.

  3. You can extract these datasets into $/your/path/to/datasets and the directory should be organized in the following structure:

    |--- /your/path/to/datasets
|       |--- BOP
|       |      |--- hope
|       |      |--- icbin
|       |      |--- icmi
|       |      |--- tless
|       |      |--- icmi
|       |      |--- lm
|       |      |--- ycbv
|       |      |--- ruapc
|       |--- RBOT
|       |--- BCOT
|       |--- OPT
|       |--- SUN2012

  4. Build the dataset symlinks

    REPO=/path/to/DeepAC
ln -s /your/path/to/datasets $REPO/datasets

Data Preprocess

DeepAC need to pre-generate contour points from various perspectives of the object, which can avoid mesh rendering duiring the tracking. You can use the following command:

sh ./script/prerender_for_training.sh
sh ./script/prerender_for_evaluation.sh

We strongly suggest utilizing the mesh.obj which is simplified in MeshLab. For example the obj in OPT dataset has million vertices and faces, but generating contour points does not need these numerous items

Training

python -m src_open.run +train=train_BOP_deepac

The results of training will be automatically saved in the REPO/workspace/train_BOP_deepac like:

|--- workspace/train_bop_deepac
|       |--- logs-2024-01-08-15-52-47
|       |         |--- train_cfg.yml
|       |         |--- model_0.ckpt
|       |         |--- model_last.ckpt
|       |--- logs-2024-01-08-19-31-15

Evaluation

  1. For RBOT, update the load_cfg and load_model of the REPO/src_open/configs/test_deepac/test_RBOT.yaml to load the trained model. Use command as following: 
    python -m src_open.run +test_deepac=test_RBOT

    The results of evaluation will be automatically saved in the REPO/workspace/test_rbot_deepac, like:

    |--- workspace/test_rbot_deepac
|       |--- logs-2024-01-08-19-09-53
|       |         |--- ape_a_regular_pose.txt # the poses of this sequence
|       |         |--- ape_a_regular.avi      # the visual result of this seqence
|       |         |--- test_results.json      # the total results
                                              # you can find the "total" in this .json about whole cm-degree and ADD metrics
|       |--- logs-2024-01-08-23-04-27

    If you want to get the results for each RBOT scene, such as "Regular", "Dynamic Light", "Noisy" and "Unmodeled Occlusion", you need to update the load_json in the REPO/src_open/configs/test_json/test_json.yaml and use the following command:

    python -m src_open.run +test_json=test_json

    The results will be printed to the terminal.

  2. For BCOT, update the load_cfg and load_model of the REPO/src_open/configs/test_deepac/test_BCOT.yaml to load the trained model. Use command as following: 
    python -m src_open.run +test_deepac=test_BCOT

    The results of evaluation will be automatically saved in the REPO/workspace/test_bcot_deepac, like:

    |--- workspace/test_bcot_deepac
|       |--- logs-2024-01-16-14-24-17
|       |         |--- 3D Touch_complex_movable_handheld_pose.txt      # the poses of this sequence
|       |         |--- 3D Touch_complex_movable_handheld_pose.avi      # the visual result of this seqence
|       |         |--- test_results.json      # the total results
                                              # you can find the "total" in this .json about whole cm-degree and ADD metrics
|       |--- logs-2024-01-08-23-04-27
  3. For OPT, use command as following for each object: 
    python -m src_open.run +test_deepac=test_OPT 'data.test_obj_names=[bike]'
python -m src_open.run +test_deepac=test_OPT 'data.test_obj_names=[chest]'
python -m src_open.run +test_deepac=test_OPT 'data.test_obj_names=[house]'
python -m src_open.run +test_deepac=test_OPT 'data.test_obj_names=[ironman]'
python -m src_open.run +test_deepac=test_OPT 'data.test_obj_names=[jet]'
python -m src_open.run +test_deepac=test_OPT 'data.test_obj_names=[soda]'

    The results of evaluation will be automatically saved in the REPO/workspace/test_opt_deepac, like: 

    |--- workspace/test_bcot_deepac
|       |--- logs-2024-01-16-14-24-17
|       |         |--- bike_fl_b_pose.txt      # the poses of this sequence
|       |         |--- bike_fl_b_pose.avi      # the visual result of this seqence
|       |         |--- test_results.json      # the total results
                                              # you can find the "total" in this .json about whole cm-degree and ADD metrics
|       |         |--- logs.txt               # you can find AUC score in this txt
|       |--- logs-2024-01-08-23-04-27

Deployment

We provide the deployment code to convert pytorch models to mlmodel via coremltools. You just need to update the the load_cfg and load_model of the REPO/src_open/configs/deploy_deepac, and then use the following command to deploy the model:

python -m src_open.run +deploy_deepac=deploy

The converted models will be automatically saved in the REPO/workspace/deploy_deepac, like:

|--- workspace/deploy_deepac
|       |--- logs-2024-01-10-15-11-06
|       |         |--- boundary_predictor.mlpackage 
|       |         |--- contour_feature_extractor0.mlpackage      
|       |         |--- contour_feature_extractor1.mlpackage      
|       |         |--- contour_feature_extractor2.mlpackage
|       |         |--- derivative_calculator.mlpackage
|       |         |--- extractor.mlpackage
|       |         |--- histogram.mlpackage

You can use these models to develop your application in Xcode. You can refer to Core ML for more information.

Citation

If you find this code useful for your research, please use the following BibTeX entry.

@inproceedings{wang2023deep,
  title={Deep Active Contours for Real-time 6-DoF Object Tracking},
  author={Long Wang and Shen Yan and Jianan Zhen and Yu Liu and Maojun Zhang and Guofeng Zhang and Xiaowei Zhou},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2023}
}

Acknowledgement

Thanks for these great optimization-based methods RBOT, 3D Object Tracking and LDT3D, which inspired us to develop this method.