You can now test our code on the provided input images in the demo folder. To this end, start the generation process for one of the config files in the configs/demo folder. For example, simply run

python generate.py configs/demo/demo_combined.yaml

This script should create a folder out/demo/demo_combined where the output meshes are stored. The script will copy the inputs into the generation/inputs folder and creates the meshes in the generation/meshes folder. Moreover, the script creates a generation/vis folder where both inputs and outputs are copied together.

Dataset

Download Datasets

To evaluate a pre-trained model or train a new model from scratch, you have to obtain the respective dataset. We use three different datasets in the DVR project:

1. ShapeNet for 2.5D supervised models (using the Choy et. al. renderings as input and our renderings as supervision)
2. ShapeNet for 2D supervised models (using the Kato et. al. renderings)
3. A subset of the DTU multi-view dataset

You can download our preprocessed data using

bash scripts/download_data.sh

and following the instructions. The sizes of the datasets are 114GB (a), 34GB (b), and 0.5GB (c).

This script should download and unpack the data automatically into the data folder.

Data Convention

Please have a look at the FAQ for details regarding the type of camera matrices we use.

Usage

When you have installed all binary dependencies and obtained the preprocessed data, you are ready to run our pre-trained models and train new models from scratch.

Generation

To generate meshes using a trained model, use

python generate.py CONFIG.yaml

where you replace CONFIG.yaml with the correct config file.

The easiest way is to use a pre-trained model. You can do this by using one of the config files which are indicated with _pretrained.yaml.

For example, for our 2.5D supervised single-view reconstruction model run

python generate.py configs/single_view_reconstruction/multi_view_supervision/ours_depth_pretrained.yaml

or for our multi-view reconstruction from RGB images and sparse depth maps for the birds object run

python generate.py configs/multi_view_reconstruction/birds/ours_depth_mvs_pretrained.yaml

Our script will automatically download the model checkpoints and run the generation. You can find the outputs in the out/.../pretrained folders.

Please note that the config files *_pretrained.yaml are only for generation, not for training new models: when these configs are used for training, the model will be trained from scratch, but during inference our code will still use the pre-trained model.

Generation From Your Own Single Images

Similar to our demo, you can easily generate 3D meshes from your own single images. To this end, create a folder which contains your own images (e.g. media/my_images). Next, you can reuse the config file configs/demo/demo_combined.yaml and just adjust the data - path and training - out_dir arguments to your needs. For example, you can set the config file to

inherit_from: configs/single_view_reconstruction/multi_view_supervision/ours_combined_pretrained.yaml
data:
  dataset_name: images
  path: media/my_images
training:
  out_dir:  out/my_3d_models

to generate 3D models for the images in media/my_images. The models will be saved to out/my_3d_models. Similar to before, to start the generation process, run

python generate.py configs/demo/demo_combined.yaml 

Note: You can only expect our model to provide reasonable results on data which is similar to what it was trained on (white background, single object, etc.).

Evaluation

For evaluation of the models, we provide the script eval_meshes.py. You can run it using

python eval_meshes.py CONFIG.yaml

The script takes the meshes generated in the previous step and evaluates them using a standardized protocol. The output will be written to .pkl/.csv files in the corresponding generation folder which can be processed using pandas.

Note: We follow previous works to use "use 1/10 times the maximal edge length of the current object’s bounding box as unit 1" (see Section 4 - Metrics). In practise, that means that we multiply the Chamfer-L1 metric by a factor of 10 for reporting the numbers in the paper.

Training

Finally, to train a new network from scratch, run

python train.py CONFIG.yaml

where you replace CONFIG.yaml with the name of the configuration file you want to use.

You can monitor on http://localhost:6006 the training process using tensorboard:

cd OUTPUT_DIR
tensorboard --logdir ./logs

where you replace OUTPUT_DIR with the respective output directory.

For available training options, please take a look at configs/default.yaml.

Futher Information

More Work on Implicit Representations

If you like the DVR project, please check out other works on implicit representions from our group:

• Mescheder et. al. - Occupancy Networks: Learning 3D Reconstruction in Function Space (CVPR 2019)
• Oechsle et. al. - Texture Fields: Learning Texture Representations in Function Space (ICCV 2019)
• Niemeyer et. al. - Occupancy Flow: 4D Reconstruction by Learning Particle Dynamics (ICCV 2019)
• Peng et. al. - Convolutional Occupancy Networks (ArXiv 2020)
• Oechsle et. al. - Learning Implicit Surface Light Fields (ArXiv 2020)

Other Relevant Works

Also check out other exciting works on inferring implicit representations without 3D supervision:

• Liu et. al. - Learning to Infer Implicit Surfaces without 3D Supervision (NeurIPS 2019)
• Sitzmann et. al. - Scene Representation Networks: Continuous 3D-Structure-Aware Neural Scene Representations (NeurIPS 2019)
• Liu. et. al. - DIST: Rendering Deep Implicit Signed Distance Function with Differentiable Sphere Tracing (CVPR 2020)
• Mildenhall et. al. - NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis (ArXiv 2020)