This repository contains the source code for applying 3PSDF to 3D reconstruction tasks based on single-view images.

This repository depends on Tensorflow, NumPy, Scikit-image and Horovod. The code is tested under the following package versions with CUDA 11.2 and Ubuntu 18.04:

tensorflow-gpu==2.6.0
numpy==1.19.5
scikit-image==0.18.3
horovod==0.23.0

• Example command with required parameters to indicate the data folders:

horovodrun -np 2 python -m src.train
  --sdf_dir data/sdf-depth7-tfrecord
  --cam_dir data/cam-tfrecord
  --img_dir data/img-tfrecord
  --split_file data/datasplit/train.lst

• Use horovodrun -np GPU_NUM to indicate the number of GPUs used for distributed training. If you only have one GPU, set GPU_NUM to 1. It is strongly recommended to use multiple GPUs.
• See python -m src.train --help for all the detailed training options.

• Example command with required parameters to indicate the data folders and pre-trained model:

python -m src.test
  --sdf_dir data/sdf-depth7-tfrecord
  --cam_dir data/cam-tfrecord
  --img_dir data/img-tfrecord
  --split_file data/datasplit/test.lst
  --load_model_path weights/3psdf_svr_weights

• See python -m src.test --help for all the detailed testing options.

To run the code:

• Train/test data (in TFRecord format): data
• Pre-trained model: weights

To obtain the raw data:

• Raw ShapeNet meshes with consistent normals: shapenet_consistent_normal.zip
• 3PSDF values sampled for ShapeNet shapes: shapnet_3psdf.zip
• Raw ShapeNet rendered images and camera parameters from 3DR2N2: shapenet_renderings.zip

To convert data:

• We provide example scripts to convert the raw data to TFRecord in src/utils/shapenet_tfrecord_generator.py

If you have any questions, please email Weikai Chen and Cheng Lin at [email protected].