Pixel2Mesh

This is an implementation of Pixel2Mesh in PyTorch. Besides, we also:

• Provide retrained Pixel2Mesh checkpoints. Besides, the pretrained tensorflow pretrained model provided in official implementation is also converted into a PyTorch checkpoint file for convenience.
• Provide a modified version of Pixel2Mesh whose backbone is ResNet instead of VGG.
• Clarify some details in previous implementation and provide a flexible training framework.

If you have any urgent question or issue, please contact [email protected].

Get Started

Environment

Current version only supports training and inference on GPU. It works well under dependencies as follows:

• Ubuntu 16.04 / 18.04
• Python 3.7
• PyTorch 1.1
• CUDA 9.0 (10.0 should also work)
• OpenCV 4.1
• Scipy 1.3
• Scikit-Image 0.15

Some minor dependencies are also needed, for which the latest version provided by conda/pip works well:

easydict, pyyaml, tensorboardx, trimesh, shapely

Two another steps to prepare the codebase:

1. git submodule update --init to get Neural Renderer ready.
2. python setup.py install in directory external/chamfer and external/neural_renderer to compile the modules.

Datasets

We use ShapeNet for model training and evaluation. The official tensorflow implementation provides a subset of ShapeNet for it, you can download it here. Extract it and link it to data_tf directory as follows. Before that, some meta files here will help you establish the folder tree, demonstrated as follows.

P.S. In case more data is needed, another larger data package of ShapeNet is also available. You can extract it and place it in the data directory. But this would take much time and needs about 300GB storage.

P.S.S. For the larger data package, we provide a temporal access here on OneDrive.

datasets/data
├── ellipsoid
│   ├── face1.obj
│   ├── face2.obj
│   ├── face3.obj
│   └── info_ellipsoid.dat
├── pretrained
│   ... (.pth files)
└── shapenet
    ├── data (larger data package, optional)
    │   ├── 02691156
    │   │   └── 3a123ae34379ea6871a70be9f12ce8b0_02.dat
    │   ├── 02828884
    │   └── ...
    ├── data_tf (standard data used in official implementation)
    │   ├── 02691156 (put the folders directly in data_tf)
    │   │   └── 10115655850468db78d106ce0a280f87
    │   ├── 02828884
    │   └── ...
    └── meta
        ...

Difference between the two versions of dataset is worth some explanation:

• data_tf has images of 137x137 resolution and four channels (RGB + alpha), 175,132 samples for training and 43,783 for evaluation.
• data has RGB images of 224x224 resolution with background set all white. It contains altogether 1,050,240 for training and evaluation.

P.S. We trained model with both datasets and evaluated on both benchmarks. To save time and align our results with the official paper/implementation, we use data_tf by default.

Usage

Configuration

You can modify configuration in a yml file for training/evaluation. It overrides dsefault settings in options.py. We provide some examples in the experiments directory.

Training

python entrypoint_train.py --name xxx --options path/to/yaml

P.S. To train on slurm clusters, we also provide settings reference. Refer to slurm folder for details.

Evaluation

python entrypoint_eval.py --name xxx --options path/to/yml --checkpoint path/to/checkpoint

Inference

You can do inference on your own images by a simple command:

python entrypoint_predict.py --options /path/to/yml --checkpoint /path/to/checkpoint --folder /path/to/images

P.S. we only support do training/evaluation/inference with GPU by default.

Results

We tested performance of some models. The official tensorflow implementation reports much higher performance than claimed in the original paper as follows. The results are listed as follows, which is close to that reported in MeshRCNN. The original paper evaluates result on simple mean, without considerations of different categories containing different number of samples, while some later papers use weighted-mean. We report results under both two metrics for caution.

Checkpoint	Eval Protocol	CD	F1τ	F12τ
Official Pretrained (tensorflow)	Mean	0.482	65.22	78.80
	Weighted-mean	0.439	66.56	80.17
Migrated Checkpoint	Mean	0.498	64.21	78.03
	Weighted-mean	0.451	65.67	79.51
ResNet	Mean	0.443	65.36	79.24
	Weighted-mean	0.411	66.13	80.13

P.S. Due to time limit, the resnet checkpoint has not been trained in detail and sufficiently.

Pretrained checkpoints

• VGG backbone: The checkpoint converted from official pretrained model (based on VGG) can be downloaded here. (scripts to migrate tensorflow checkpoints into .pth are available in utils/migrations. )
• ResNet backbone: As we provide another backbone choice of resenet, we also provide a corresponding checkpoint here.

Details of Improvement

We explain some improvement of this version of implementation compared with the official version here.

• Larger batch size: We support larger batch size on multiple GPUs for training. Since Chamfer distances cannot be calculated if samples in a batch with different ground-truth pointcloud size, "resizing" the pointcloud is necessary. Instead of resampling points, we simply upsample/downsample from the dataset.
• Better backbone: We enable replacing VGG by ResNet50 for model backbone. The training progress is more stable and final performance is higher.
• More stable training: We do normalization on the deformed sphere, so that it's deformed at location $(0,0,0)$; we use a threshold activation on $z$-axis during projection, so that $z$ will always be positive or negative and never be $0$. These seem not to result in better performance but more stable training loss.

Demo

Generated mesh samples are provided in datasets/examples from our ResNet model. Three mesh models in a line are deformed from a single ellipsoid mesh with different number of vertices (156 vs 268 vs 2466) as configurated in the original paper.

Acknowledgements

Our work is based on the official version of Pixel2Mesh; Some part of code are borrowed from a previous PyTorch implementation of Pixel2Mesh. The packed files for two version of datasets are also provided by them two. Most codework is done by Yuge Zhang.