Awesome

PoseContrast

[3DV 2021 Oral] Pytorch implementation of Class-Agnostic Object Viewpoint Estimation in the Wild with Pose-Aware Contrastive Learning paper.

Check out our Paper and Webpage for more details.

@INPROCEEDINGS{Xiao2020PoseContrast,
    author    = {Yang Xiao and Yuming Du and Renaud Marlet},
    title     = {PoseContrast: Class-Agnostic Object Viewpoint Estimation in the Wild with Pose-Aware Contrastive Learning},
    booktitle = {International Conference on 3D Vision (3DV)},
    year      = {2021}
}

Installation

1. Create conda environment

conda env create -f environment.yml
conda activate PoseContrast

PyTorch version: This repo has been tested under PyTorch 1.0.0 and PyTorch 1.6.0 with similar performances.

2. Download datasets

cd ./data
bash download_data.sh

This command will download following datasets:

Pascal3D+ (link to the original dataset page. ImageNet and PascalVOC images picturing 12 rigid object classes in the wild)
Pix3D (link to the original dataset page. Mixed-sources images picturing 9 rigid object classes including "tools" and "misc".)
ObjectNet3D (link to the original dataset page. ImageNet images picturing 100 rigid object classes.)

Note: Pix3D is only used for evaluation, while Pascal3D+ and ObjectNet3D both contain train/val split for training and testing.

3. Download pretrained models

cd ./pretrain_models
bash download_moco.sh

This command will download the MOCOv2 model released by FAIR and correct the module names.

mkdir exps/PoseContrast_Pascal3D_MOCOv2 && cd exps/PoseContrast_Pascal3D_MOCOv2
wget https://www.dropbox.com/s/mlmubnz9xgbflm4/ckpt.pth?dl=0

This command will download our class-agnostic object viewpoint estimation network trained on Pascal3D+ dataset.

The model trained on ObjectNet3D dataset can be downloaded from GoogleDrive.

PoseContrast_ObjectNet3D_ZeroShot contains the trained model for the first base-training stage, where the model is trained only on the 80 base classes.
PoseContrast_ObjectNet3D_FewShot contains the trained model for the second fine-tuning stage, where the previous model is fine-tuned on both the 80 base classes and the 20 novel classes.

How to use

1. Launch a training

./scripts/train.sh

Training on the Pascal3D+ dataset. Models are saved at exps/PoseContrast_Pascal3D_MOCOv2, a training log file trainer.log will also be generated.

./scripts/train_object3d.sh

Training on the ObjectNet3D dataset. Models are saved at exps/PoseContrast_ObjectNet3D_ZeroShot and PoseContrast_ObjectNet3D_FewShot, respectively.

2. Evaluate on different datasets

./scripts/test.sh

Evaluate the model on Pascal3D+ and Pix3D:

prediction folder would be created to save the predicted viewpoints,
correlation folder would be created to save the angle classification scores and viewpoint estimation errors,
testing log file tester.log would be generated saving the quantitative evaluation results.

./scripts/test_object3d.sh

Evaluate the model on the ObjectNet3D dataset and report the results for base and novel classes.

Visual Results

Further information

If you like this project, please check out related works on object pose estimation from our group: