The 3D object is transformed into a 56×56×56 voxel grid which is the input size for our model. The entropy model uses a branching architecturethat converges into a fully connected layer. The output of the entropy model is a 60-values vector which is reshaped into a 2D entropy map. From the entropymap a peak selection algorithm returns the coordinates of the local maxima. We extract the views corresponding to those coordinates and we use them asinput for as many CNNs. The label and pose predictions are pooled by majority vote, resulting in the final prediction.
Download and unzip the Modelet10 dataset from the official distribution: http://3dvision.princeton.edu/projects/2014/3DShapeNets/ModelNet10.zip (451MB)
Ensure you are running Python>=3.6.5 and import the required libraries by running:
pip install -r requirements.txt
WARNING: The dataset is quite large and generation scripts may take hours to complete.
To generate the entropy dataset run:
python generate_entropy_dataset.py --modelnet10 <dir> --out <out>
substitute <dir> with the root directory of the ModelNet10 dataset.
This will save the entropy values for each of the 60 depth-views of
each object in the dataset in the file <out>/entropy_dataset.csv.
You can change the <out> directory with the --out parameter.
To generate the view dataset run:
python generate_view_dataset.py --modelnet10 <dir> --set train --out <out>
python generate_view_dataset.py --modelnet10 <dir> --set test --out <out>
substitute <dir> with the root directory of the ModelNet10 dataset.
The output will consist in the train and test set of the image-views
stored respectively in <out>/view-dataset-train/image and
<out>/view-dataset-test/image. You can change the <out> directory with the
--out parameter.
The entropy model requires a pre-vozelized dataset which is generated by running:
python prevoxelization.py --modelnet10 <dir>
As usual substitute <dir> with the root directory of the ModelNet10 dataset.
The dataset will be generated in the <out>/voxel_data folder. You can change the
<out> directory with the --out parameter.
The entropy model requires 50x50x50 occupancy grids in numpy format which
are the results of prevoxelization.py, and to parse the entropy data
from the entropy_dataset.csv file:
python entropy_model.py --voxel_data <dir> --entropy_dataset <csv> --epochs <e> --batch_size <b> --out <out>
Substitute <dir> with /voxel_data and <csv> with
entropy-dataset/entropy_dataset.csv (or with the output directory you choose).
You can select the number of epochs and the batch size by substituting the corresponding
parameter.
The model with the best validation loss will be stored at
<out>/entropy-model/entropy_model.h5 alongside the logs of the training.
The classification model requires to be trained on the dataset /view-dataset-train/image generated,
using /view-dataset-test/image as validation data, generated with generate_view_dataset.py:
python single_view_cnn.py \
--train_data=view-dataset-train/image/ \
--test_data=view-dataset-test/image/ \
--batch_size=<b> \
--epochs=<e> \
--architecture=<arch> \
--train_sample_rate=1 \
--test_sample_rate=50 \
--lr=<lr>
Substitute <b>, <e> and <lr>with the batch size, the number of epochs
and the learning rate to train the model.
The architecture can be specified by substituting <arch> with vgg or
mobilenetv2.
The train/test_sample_ratio are parameters to train the model on a 1:r
subset of the full dataset.
Due to the large size of the dataset we suggest to reduce the validation
data to a 1:50 ratio and to expect at least 1.5h per epoch of training time
for a 1:1 train set. To achieve comparable results while reducing substantially
training time use a higher subsampling ratio and increase the learning rate of at
least an order of magnitude.
Pre-trained models are available here:
entropy_model.h5
classifier_model.h5
A demonstration of the classification from a 3D mesh file can be run as following:
python demo.py --data <off_file> --entropy_model <e_model> --classifier_model <c_model>
Substitute <off_file> with an object from the ModelNet10 dataset, <e_model> with the
model trained with entropy_model.py and <c_model> with the model trained with single_view_cnn.py.
The demo will show the color coded entropy map extracted by the entropy model with
black circles marking the selected best-views. Then it will render and plot the images
corresponding to the selected views.
On terminal it will report the extracted views with their view-code [0 - 59] and it will
show the majority voting for class and pose.
Latest version available on arXiv (March 2021)
Please adequately refer to the papers any time this code is being used. If you do publish a paper where MORE helped your research, we encourage you to cite the following paper in your publications:
@article{MREE2021,
title={MORE: Simultaneous Multi-View 3D Object Recognition and Pose Estimation},
author={Parisotto Tommaso and Kasaei, Hamidreza},
journal={arXiv preprint arXiv:3655085},
year={2021}
}
Tommaso Parisotto and Hamidreza Kasaei
Work done while at RUG.