ZeroShape

[Project Page] [Paper] [Demo]

The repository currently includes demo, training, evaluation code and data for ZeroShape.

Dependencies

If your GPU supports CUDA 10.2, please install the dependencies by running

conda env create --file requirements.yaml

If you need newer CUDA versions, please install the dependencies manually:

conda create -n zeroshape python=3 pytorch::pytorch=1.11 pytorch::torchvision=0.12 cudatoolkit=10.2 (change these to your desired version)
conda install -c conda-forge tqdm pyyaml pip matplotlib trimesh tensorboard
pip install pyrender opencv-python pymcubes ninja timm

To use the segmentation preprocessing tool, please run

pip install rembg

Demo

Please download the pretrained weights for shape reconstruction at this url and place it under the weights folder. We have prepared some images and masks under the examples folder. To reconstruct their shape, please run:

python demo.py --yaml=options/shape.yaml --task=shape --datadir=examples --eval.vox_res=128 --ckpt=weights/shape.ckpt

The results will be saved under the examples/preds folder.

To run the demo on your own images and masks, feel free to drop them in the examples folder. If you do not have mask, please run:

python preprocess.py path-to-your-image

The preprocessed image and mask will be saved in the my_examples folder. To reconstruct their shape, please run:

python demo.py --yaml=options/shape.yaml --task=shape --datadir=my_examples --eval.vox_res=128 --ckpt=weights/shape.ckpt

If you want to estimate the visible surface (depth and intrinsics), please download the pretrained weights for visible surface estimation at this url and place it under the weights folder. Then run:

python demo.py --yaml=options/depth.yaml --task=depth --datadir=examples --ckpt=weights/depth.ckpt

The results will be saved under the examples/preds folder.

Data

Please download our curated training and evaluation data at the following links:

After extracting the data, organize your data folder as follows:

data
├── train_data/
|   ├── objaverse_LVIS/
|   |   ├── images_processed/
|   |   ├── lists/
|   |   ├── ...
|   ├── ShapeNet55/
|   |   ├── images_processed/
|   |   ├── lists/
|   |   ├── ...
├── OmniObject3D/
|   ├── images_processed/
|   ├── lists/
|   ├── ...
├── Ocrtoc/
|   ├── images_processed/
|   ├── lists/
|   ├── ...
├── Pix3D/
|   ├── img_processed/
|   ├── lists/
|   ├── ...
├── ...

Note that you do not have to download all the data. For example, if you only want to perform evaluation on one of the data source, feel free to only download and organize that specific one accordingly.

Training

The first step of training ZeroShape is to pretrain the depth and intrinsics estimator. If you have downloaded the weights already (see demo), you can skip this step and use our pretrained weights at weights/depth.ckpt. If you want to train everything from scratch yourself, please run

python train.py --yaml=options/depth.yaml --name=run-depth

The visualization and results will be saved at output/depth/run-depth. Once the training is finished, copy the weights from output/depth/run-depth/best.ckpt to weights/depth.ckpt.

To train the full reconstruction model, please run

python train.py --yaml=options/shape.yaml --name=run-shape

The visualization and results will be saved at output/shape/run-shape.

Evaluating

To evaluate the model on a specific test set (omniobj3d|ocrtoc|pix3d), please run

python evaluate.py --yaml=options/shape.yaml --name=run-shape --data.dataset_test=name_of_test_set --eval.vox_res=128 --eval.brute_force --eval.batch_size=1 --resume

The evaluation results will be printed and saved at output/depth/run-shape. If you want to evaluate the checkpoint we provided instead, feel free to create an empty folder output/shape/run-shape and move weights/shape.ckpt to output/shape/run-shape/best.ckpt

References

If you find our work helpful, please consider citing our paper.

@article{huang2023zeroshape,
  title={ZeroShape: Regression-based Zero-shot Shape Reconstruction},
  author={Huang, Zixuan and Stojanov, Stefan and Thai, Anh and Jampani, Varun and Rehg, James M},
  journal={arXiv preprint arXiv:2312.14198},
  year={2023}
}