Skip to content

An autoregressive model for point cloud generation augmented with self-attention

Notifications You must be signed in to change notification settings

Xinpeng-Wang/PointGrow

 
 

Repository files navigation

PointGrow: Autoregressively Learned Point Cloud Generation with Self-Attention

This work presents a novel autoregressive model, PointGrow, which generates realistic point cloud samples from scratch or conditioned on given semantic contexts. This model operates recurrently, with each point sampled according to a conditional distribution given its previously-generated points. It is further augmented with dedicated self-attention modules to capture long-range interpoint dependencies during the generation process.

[Project] [Paper]

Data

We provided processed point clouds from 7 categories in ShapeNet, including airplane, car, table, chair, bench, cabinet and lamp. The coordinates of those point clouds, arranged as (z, y, x), range from 0 to 1. They are sorted in the order of z, y and x, and can be downloaded from here.

Unconditional PointGrow

The model is trained per category, change the ShapeNet category id when working on different categories.

Category Id
Airplane      02691156
Car           02958343
Table 04379243
Chair         03001627
Bench         02828884
Cabinet       02933112
Lamp 03636649
  • Run unconditional PointGrow training script for airplane category with SACA-A module:
python train_unconditional.py --cat 02691156 --model unconditional_model_saca_a

Model parameters will be stored under "log/unconditional_model_saca_a/02691156".

  • For example, to generate 300 point clouds for airplane category using the pre-trained model:
python generate_unconditional.py --cat 02691156 --model unconditional_model_saca_a --tot_pc 300

The generated point clouds will be stored in the format of numpy array under "res/unconditional_model_saca_a/res_02691156.npy".

Conditional PointGrow

One-hot categorical vectors

Generate point clouds conditioned on additional one-hot vectors, with their non-empty elements indicating ShapeNet categories. For example, following the order in the above table, the one-hot vector for airplane can be expressed as [1, 0, 0, 0, 0, 0, 0].

  • Train conditional PointGrow with one-hot vectors:
python train_conditional_one_hot.py

Model parameters will be saved under "log/conditional_model_one_hot".

  • For example, to generate 50 point clouds for airplane (cat_idx = 0) using the pre-trained model:
python generate_conditional_one_hot.py --cat_idx 0 --tot_pc 50

The generated point clouds will be stored in the format of numpy array under "res/conditional_model_one_hot/res_02691156.npy".

Image embeddings

Generate point clouds conditioned on the embedding vectors of given 2D images. We still use ShapeNet point clouds, and obtain their 2D renderings from 3D-R2N2. A collection of 2D images of airplane and car categories and their shape ids matching the point clouds provided in this project can be found here.

  • Train conditional PointGrow with 2D image embeddings for airplane category:
python train_conditional_im.py --cat 02691156

Model parameters will be saved under "log/conditional_model_im/02691156".

  • For example, to generate 50 point clouds for ShapeNet airplane testing images using the pre-trained model:
python generate_conditional_im.py --tot_pc 50 --batch_size 25 --cat 02691156

The batch_size variable is recommened to be set less than 25 to fit GPU memory. The tot_pc variable will be truncated to a multiple of batch_size if tot_pc is larger than batch_size. The generated point clouds will be stored in the format of numpy array under "res/conditional_model_im/res_02691156.npy".

Citation

Please cite this paper if you want to use it in your work,

@inproceedings{sun2020pointgrow,
  title={Pointgrow: Autoregressively learned point cloud generation with self-attention},
  author={Sun, Yongbin and Wang, Yue and Liu, Ziwei and Siegel, Joshua and Sarma, Sanjay},
  booktitle={The IEEE Winter Conference on Applications of Computer Vision},
  pages={61--70},
  year={2020}
}

License

MIT License

About

An autoregressive model for point cloud generation augmented with self-attention

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%