fairuzsafwan/deepmerge

deepMerge: A model to reconstruct 3D model from depth map

deepMerge

deepMerge: A model to reconstruct 3D model from depth map by utilizing local and generic features

Requirements

• Torch
• nn
• nngraph
• cutorch
• cunn
• paths
• image
• optim
• gnuplot
• openCV

Dataset & Pretrained model

Download the Dataset and pretrained model

Setting up

After downloading the Dataset and pretrained model, unzip the files and store the files in their respective repository.

Dataset: Download the Dataset and store the folder in the Data/nonbenchmark_ownCamPos folder

The files and folder should be in the following format:

Data/nonbenchmark_ownCamPos/Datasets/test/Test-224x224-0.data
Data/nonbenchmark_ownCamPos/Datasets/test/Test-224x224-1.data
Data/nonbenchmark_ownCamPos/Datasets/test/Test-224x224-2.data

Data/nonbenchmark_ownCamPos/Datasets/train/Train-224x224-0.data
Data/nonbenchmark_ownCamPos/Datasets/train/Train-224x224-1.data
Data/nonbenchmark_ownCamPos/Datasets/train/Train-224x224-2.data
.
.
.
Data/nonbenchmark_ownCamPos/Datasets/train/Train-224x224-67.data

Data/nonbenchmark_ownCamPos/Datasets/validation/Valid-224x224-0.data
Data/nonbenchmark_ownCamPos/Datasets/validation/Valid-224x224-1.data
Data/nonbenchmark_ownCamPos/Datasets/validation/Valid-224x224-2.data
Data/nonbenchmark_ownCamPos/Datasets/validation/Valid-224x224-3.data

Pretrained model: Download the pretrained_model for epoch 80, 90 and 100 and store the folders in the pretrained_model/model repository.

The files and folder should be in the following format:

pretrained_model/model/epoch80/mean_logvar.t7
pretrained_model/model/epoch80/model.t7

pretrained_model/model/epoch90/mean_logvar.t7
pretrained_model/model/epoch90/model.t7

pretrained_model/model/epoch100/mean_logvar.t7
pretrained_model/model/epoch100/model.t7

How to run

In order to run the commands below, you need to download the Dataset and pretrained model and store them in their respective repositories

1. Using the pretrained model

run the command below:

• modelName = the name of the parent folder where the model resides (eg. pretrained_model)
• fromEpoch = select an epoch from which the model should be used for reconstructions (eg. 80)
• GPU = select a GPU to use from 0 to N (where N is the total number of GPUs available minus 1). set it to 0 if you only have one GPU (eg. 0)

sh reconstruct.sh modelName fromEpoch GPU

sh reconstruct.sh pretrained_model 80 0

2. Training the model from scratch

run the command below:

• modelName = the name of the folder where the model will reside (eg. sampleModel)
• GPU = select a GPU to use from 0 to N (where N is the total number of GPUs available minus 1). set it to 0 if you only have one GPU (eg. 0)

sh train.sh modelName GPU

sh train.sh sampleModel 0

Computing IoU (Windows only)

1. Install the following requirements:

• Visual C++ Redistributable for Visual Studio 2015
• Visual C++ Redistributable Packages for Visual Studio 2013
• python
• Anaconda
• openCV

2. Register openCV path in environment variables

C:\opencv\build\\x64\vc14\bin
C:\opencv\build\\x64\vc15\bin

3. run Command Prompt and cd to Compute_IoU

cd Compute_IoU

4. cd to Compute_IoU

copy epoch folder where all 57 categories of depth maps and silhouettes are reconstructed into input/ folder

cp -r deepMerge\pretrained_model\experiments\epoch80 C:\Users\safwan\Desktop\Compute_IoU\input\

5. How to run

run the command below:

• experimentName = the name of the experiment folder where the 3D models will be reconstructed (eg. experimentSample)

compute_IoU.bat experimentName

compute_IoU.bat experimentSample

Output: Reconstructed models will reside in output/ folder