NeMo

This is the repo for the series works on Neural Mesh Models. In this repo, we implement 3D object pose estimation, 3D object pose estimation via VoGE renderer, 6D pose object estimation, object classification, and cross domain training. The original implementation of NeMo is here.

Release Note on Sept 10 (by Angtian)

Introduce a major refactor, main for feature banks and "mask remove" functions. In the new implementation, the feature banks support multiple objects in each training image with different class labels. Note the implementation of multiple classes is CUDA-based, which requires installation of the specific CUDA layer. (Running 3D pose only should be fine without these CUDA layers.) To install:

cd cu_layers
python setup.py install

After the installation, you will find a lib named "CuNeMo" in your Python libs. Previous configs should be compatible except for changes in config/model

memory_bank:
    class_name: nemo.models.feature_banks.FeatureBankNeMo

The previous implementation of classification NeMo is removed, we will add support for classification NeMo very soon. Contact me directly if you find any bugs or compatibility issues.

Features

Easily train and evaluate neural mesh models for multiple tasks:

• 3D pose estimation
• 6D poes estimation
• 3D-aware image classification
• Amodal segmenation

Experiment on various benchmark datasets:

• PASCAL3D+
• Occluded PASCAL3D+
• ObjectNet3D
• OOD-CV
• SyntheticPASCAL3D+

Reproduce baseline models for fair comparison:

• Regression-based models (ResNet50, Faster R-CNN, etc.)
• Transformers
• StarMap

Installation

Environment (manual setup)

1. Create conda environment:

conda create -n nemo python=3.9
conda activate nemo

2. Install PyTorch (see pytorch.org):

conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=10.2 -c pytorch

3. Install PyTorch3D (see github.com/facebookresearch/pytorch3d):

conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install -c bottler nvidiacub
conda install pytorch3d -c pytorch3d

4. Install other dependencies:

conda install numpy matplotlib scipy scikit-image
conda install pillow
conda install -c conda-forge timm tqdm pyyaml transformers
pip install git+https://github.com/NVlabs/nvdiffrast/
pip install wget gdown BboxTools opencv-python xatlas pycocotools seaborn wandb

3. (Optional) Install VoGE (see github.com/Angtian/VoGE):

pip install git+https://github.com/Angtian/VoGE.git

Environment (from yml)

In case the previous method failed, setup the environment from a compiled list of packages:

conda env create -f environment.yml
pip install git+https://github.com/NVlabs/nvdiffrast/
pip install -e .

Data Preparation

See data/README.

Quick Start

Train and evaluate a neural mesh model (NeMo) on PASCAL3D+ for 3D pose estimation:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 scripts/train.py \
    --cate car \
    --config config/omni_nemo_pose_3d.yaml \
    --save_dir exp/pose_estimation_3d_nemo_car

CUDA_VISIBLE_DEVICES=0 python3 scripts/inference.py \
    --cate car \
    --config config/omni_nemo_pose_3d.yaml \
    --save_dir exp/pose_estimation_3d_nemo_car \
    --checkpoint exp/pose_estimation_3d_nemo_car/ckpts/model_800.pth

NeMo with VoGE:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 scripts/train.py \
    --cate car \
    --config config/omni_voge_pose_3d.yaml \
    --save_dir exp/pose_estimation_3d_voge_car

CUDA_VISIBLE_DEVICES=0 python3 scripts/inference.py \
    --cate car \
    --config config/omni_voge_pose_3d.yaml \
    --save_dir exp/pose_estimation_3d_voge_car \
    --checkpoint exp/pose_estimation_3d_voge_car/ckpts/model_800.pth

NeMo on PASCAL3D+ without scaling during data pre-processing:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 scripts/train.py \
    --cate car \
    --config config/omni_nemo_pose_3d_ori.yaml \
    --save_dir exp/pose_estimation_3d_ori_car

CUDA_VISIBLE_DEVICES=0 python3 scripts/inference.py \
    --cate car \
    --config config/omni_nemo_pose_3d_ori.yaml \
    --save_dir exp/pose_estimation_3d_ori_car \
    --checkpoint exp/pose_estimation_3d_ori_car/ckpts/model_800.pth

Train and evaluate a regression-based model (ResNet50-General) on PASCAL3D+ for 3D pose estimation:

CUDA_VISIBLE_DEVICES=0 python3 scripts/train.py \
    --cate all \
    --config config/pose_estimation_3d_resnet50_general.yaml \
    --save_dir exp/pose_estimation_3d_resnet50_general_car

CUDA_VISIBLE_DEVICES=0 python3 scripts/inference.py \
    --cate car \
    --config config/pose_estimation_3d_resnet50_general.yaml \
    --save_dir exp/pose_estimation_3d_resnet50_general \
    --checkpoint exp/pose_estimation_3d_resnet50_general/ckpts/model_90.pth

Pre-trained Models

Pre-trained Models for 3D pose estimation

The pre-trained model for NeMo model:

https://drive.google.com/file/d/14fByOZs_Zzd-97Ulk2BKJhVNFKAnFWvg/view?usp=sharing

3D pose	plane	bike	boat	bottle	bus	car	chair	table	mbike	sofa	train	tv	Mean
Pi/6 Pi/18 Med	86.9 55.3 8.94	80.3 30.9 15.51	77.4 50.2 9.95	90.0 56.9 8.24	95.3 91.5 2.66	98.9 96.5 2.71	89.1 56.7 8.68	80.2 63.1 6.96	86.6 33.2 13.34	95.8 65.9 7.18	64.4 55.3 7.32	82.0 48.6 10.61	87.4 65.5 7.42

The pre-trained model for NeMo-VoGE model:

https://drive.google.com/file/d/1kogFdjVbOIuSlKx1NQ1c1XEjbvJEQWJg/view?usp=sharing

3D pose	plane	bike	boat	bottle	bus	car	chair	table	mbike	sofa	train	tv	Mean
Pi/6 Pi/18 Med	87.8 62.3 7.57	82.9 36.7 14.02	75.4 51.0 9.7	88.2 55.2 9.1	97.4 94.5 2.38	99.0 96.4 2.89	90.7 54.9 8.96	83.6 69.7 5.7	87.4 39.1 12.3	94.4 65.4 7.77	91.3 83.3 3.84	80.5 54.4 8.80	89.5 69.5 6.82

The pre-trained model for NeMo model without scaling:

https://drive.google.com/file/d/1ybVTDx6DvV_H01SUZkKqWQjKu-BfweGJ/view?usp=sharing

3D pose	plane	bike	boat	bottle	bus	car	chair	table	mbike	sofa	train	tv	Mean
Pi/6 Pi/18 Med	83.0 48.0 10.62	75.7 24.7 18.54	68.3 34.0 14.97	84.5 44.3 11.67	96.2 90.0 3.00	98.8 95.4 3.12	85.8 44.6 11.01	80.4 58.5 8.07	78.1 26.6 15.22	94.6 58.8 8.31	79.2 64.0 6.65	85.8 45.6 11.25	86.0 60.2 8.99

The pre-trained model for NeMo-VoGE model without scaling:

https://drive.google.com/file/d/10ggpneADVWClXWx42yQeJ_unFt53oQ1I/view?usp=sharing

3D pose	plane	bike	boat	bottle	bus	car	chair	table	mbike	sofa	train	tv	Mean
Pi/6 Pi/18 Med	83.1 51.9 9.56	80.2 29.9 16.33	68.1 36.3 14.97	83.9 44.6 11.07	98.1 94.2 2.92	98.3 93.2 3.75	89.0 50.1 9.97	83.0 65.0 6.70	81.8 32.8 14.06	94.1 61.4 8.03	90.5 76.1 5.45	83.7 46.4 10.70	87.4 62.9 8.51

Documentation

See documentation.

Citation

@inproceedings{wang2021nemo,
   title={NeMo: Neural Mesh Models of Contrastive Features for Robust 3D Pose Estimation},
   author={Angtian Wang and Adam Kortylewski and Alan Yuille},
   booktitle={International Conference on Learning Representations},
   year={2021},
   url={https://openreview.net/forum?id=pmj131uIL9H}
}
@software{nemo_code_2022,
   title={Neural Mesh Models for 3D Reasoning},
   author={Ma, Wufei and Jesslen, Artur and Wang, Angtian},
   month={12},
   year={2022},
   url={https://github.com/wufeim/NeMo},
   version={1.0.0}
}

Further Information

This repo builds upon several previous works:

• NeMo: Neural Mesh Models of Contrastive Features for Robust 3D Pose Estimation (ICLR 2021)
• Robust Category-Level 6D Pose Estimation with Coarse-to-Fine Rendering of Neural Features (ECCV 2022)

Acknowledgements

In this project, we borrow codes from several other repos:

• NeMo by Angtian Wang in Angtian/NeMo
• DMTet by NVIDIA in nv-tlabs/GET3D
• torch_utils by NVIDIA in nv-tlabs/GET3D
• uni_rep by NVIDIA in nv-tlabs/GET3D
• dnnlib by NVIDIA in nv-tlabs/GET3D