cnsystem / xnocs

Multiview Aggregation for Learning Category-Specific Shape Reconstruction, NeurIPS 2019

Multiview Aggregation for Learning Category-Specific Shape Reconstruction

NeurIPS 2019

Srinath Sridhar¹, Davis Rempe¹, Julien Valentin², Sofien Bouaziz², Leonidas J. Guibas^1,3
1Stanford University, ²Google Inc., ³Facebook AI Research

About

This repository contains the code and datasets for the NeurIPS 2019 paper Multiview Aggregation for Learning Category-Specific Shape Reconstruction. If you use the code or datasets please cite our paper.

@InProceedings{xnocs_sridhar2019,
author = {Sridhar, Srinath and Rempe, Davis and Valentin, Julien and Bouaziz, Sofien and Guibas, Leonidas J.},
title = {Multiview Aggregation for Learning Category-Specific Shape Reconstruction},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2019}
}

Code

1. Download and install the following required packages. We only support Python 3.5+.

• PyTorch 1.0+ (see the PyTorch website for installation instructions)
• numpy>=1.15.4
• opencv-python<=3.2.0.8
• matplotlib
• palettable (pip install palettable)
• tk3dv (pip install git+https://github.com/drsrinathsridhar/tk3dv.git)

2. Download the datasets (see below for details)

Note: The training code can automatically download the ShapeNetCOCO dataset. Please see the dataset readme for more details.

• shapenetcoco_v1.zip (172 GB)
• shapenetplain_v1.zip (5 GB)

3. Download the trained weights

These are weights trained on the ShapeNetCOCO dataset and reported in Table 2 in the paper.

• Single View Network (1 GB)
• Multi-view Network (upto 5 views) with Permutation Equivariant layers (1 GB)

4. Sample code

The main training and evaluation code is noxray/nxm/nxm.py for the single-view network and noxray/mv_nxm.py

To train the single-view network on the cars category in ShapeNetCOCO, use this command:

python nxm.py --mode train --input-dir <BASE DATASET DIR> --output-dir <BASE OUTPUT DIR> --expt-name XNOCS_SV --category cars --arch SegNetSkip --seed 0 --data-limit <SKIP THIS TO USE FULL DATA> --val-data-limit <SKIP THIS TO USE FULL DATA>

You can also load the above config from a file using the @ seperator. Note: in the paper we use --learning-rate 0.0001 and --batch-size 1. The code is not well-tested with larger batch sizes and there have been reported issues with doing so. For reference, when using the ShapeNetPlain dataset in the paper we train for --epochs 100 which takes just over 1 day for both the Cars and Airplanes categories (NVIDIA V100 GPU). Training on ShapeNetCOCO is considerably slower.

To evaluate the network and write the output images, use this command:

python nxm.py --mode val --input-dir <BASE DATASET DIR> --output-dir <BASE OUTPUT DIR> --expt-name XNOCS_SV --category cars --arch SegNetSkip --seed 0 --test-samples 1000000

Note: the batch size is currently hard-coded to 1 for evaluation.

The multi-view network commands are very similar to the above.

Datasets

We provide two datasets: ShapeNetCOCO and ShapeNetPlain. ShapeNetCOCO is the primary dataset of which we demonstrate reconstruction results. ShapeNetPlain is the dataset we use for comparison with 3D-R2N2 and DPC.

• shapenetcoco_v1.zip (172 GB)
• shapenetplain_v1.zip (5 GB)

Please see the dataset readme for more details.

Contact

Srinath Sridhar ([email protected])