Home

Awesome

Learning Single-View 3D Reconstruction with Limited Pose Supervision

TensorFlow implementation for the paper:

Learning Single-View 3D Reconstruction with Limited Pose Supervision

Guandao Yang, Yin Cui, Serge Belongie, Bharath Hariharan

Dependency

The recommended way to install the dependency is

pip install -r requirements.txt

Preparation

Please use the following Google Drive link to download the datasets: [drive]. There are two files : data.tar.gz and ShapeNetVox32.tar.gz. Please download both of them and uncompressed it into the project root directory:

tar -xvf ShapeNetVox32.tar.gz
tar -xvf data.tar.gz
rm ShapeNetVox32.tar.gz
rm data.tar.gz

Training

In order to train a model, please use train.py script. The default hyper-parameter are stored in config.py, and all the training settings are stored in training_settings.py. For example, to train single category chairs with 50% pose annotations, we could use:

python train.py --data chairs_pose0.5

where chairs_pose0.5 refers to an entry in the training_settings.py file.

The scripts folder contains command line arguments for running different experiments. Note that each script contains multiple training commands.

For example, in order to run all the single category training with 50% of pose annotations, please use:

./scripts/single_category.sh

To train all multi-category models or to pre-train all models for few-shot transfer learning, please use:

./scripts/muti.sh

Evaluation

File inference.py contains codes to load a trained model and evaluate on specific data split or categories. scripts/get_score.py helps to organize the evaluation scores into .csv files. For Detail usage, please refer to scripts/eval_*.sh.

For example, if you want to evaluate all the single category training experiments, try running

./sripts/eval_single_category.sh

Results

Following results are reported from the testing set. Please compare to these results during reproduction.

Single Category Experiments

SettingMaxIoUAPIoU(t=0.4)Iou(t=0.5)
airplanes_pose0.50.4840.63770.43960.4235
beches_pose0.50.3680.48220.34210.3379
cars_pose0.50.74210.86980.69760.6777
chairs_pose0.50.44590.57030.41130.3896
sofas_pose0.50.55230.69540.52760.5224
tables_pose0.50.41710.55410.3550.3402

Multiple Category (AP)

CategoryPose 100%Pose 50%Pose 10%Pose 1%
airplanes0.71030.70620.64570.5316
cars0.92230.91390.88880.797
chairs0.59220.57380.53250.4079
displays0.60250.59170.46940.2857
phones0.82940.8130.6620.498
speakers0.70350.68690.63360.5481
tables0.56030.54860.48270.3948
Mean0.70290.69060.61640.4947

Out of category (AP)

CategoryPose 100%Pose 50%Pose 10%Pose 1%
benches0.42430.40440.33910.2485
cabinets0.63130.61230.56670.5321
vessels0.61090.60630.55810.5325

TODO

We will release the codes for few-shot experiments soon.

Citation

If you find this our works helpful for your research, please cite:

@InProceedings{Yang_2018_ECCV,
author = {Yang, Guandao and Cui, Yin and Belongie, Serge and Hariharan, Bharath},
title = {Learning Single-View 3D Reconstruction with Limited Pose Supervision},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {September},
year = {2018}
}