Skip to content

Latest commit

 

History

History
164 lines (138 loc) · 6.61 KB

README.md

File metadata and controls

164 lines (138 loc) · 6.61 KB
Raw

MURAUER: Mapping Unlabeled Real Data for Label AUstERity

This repository contains the code for the semi-supervised method introduced in:

MURAUER: Mapping Unlabeled Real Data for Label AUstERity
Georg Poier, Michael Opitz, David Schinagl and Horst Bischof.
In Proc. WACV, 2019. (Project Page).

Architecture sketch for the introduced system
We learn to map from the features of real data to the features of synthetic data. The mapping is mainly learned from a large amount of synthetic as well as unlabeled real data. We exploit unlabeled data using two auxiliary objectives, which enforce that (i) the mapped representation is pose specific and (ii) at the same time, the distributions of real and synthetic data are aligned. While the distributions are aligned using an adversarial term, pose specificity is enforced by building on our previous work. By exploiting unlabeled data in this way we can significantly improve the results for hand pose estimation and outperform many recent approaches already with about 1% of the labeled real data they use.

Usage

  1. Download dataset
  2. Adapt paths in configuration to point to the dataset
  3. Run code

Download dataset

We provide a data-loader the NYU dataset [1].

Adapt configuration

You need to change the respective paths (to the location of the dataset) in config/config_data_nyu.py.

Run code

python main_run.py

By default, this will first pre-train the model using only synthetic data, then train the model using all data, and finally test it (see below for using pre-trained models). The code will log the training and validation error using crayon (see below), and output final results in the results folder. To change further settings you can adapt the respective configuration files in the config folder or via the command-line (see python main_run.py --help for details). The default settings should reproduce the results from the paper (actually the results should be slightly better than those from the paper - mainly due to a bug fix wrt. some non-linearities and training with a newer pytorch version (1.0.1)). Results using the pre-trained models:

n 100 1,000 10,000 72,757
Mean error (mm) 11.91 9.28 9.24 9.19

Use pre-trained model

We provide a model pre-trained only on synthetic data (to skip the pre-training) as well as fully trained models (using different numbers of labeled real samples).

Use model pre-trained with synthetic data

Download the model (filename: model_synth_only.mdl_pretrain) from our server. To use the default configuration, place the model in results/pretrained_models:

mkdir results/pretrained_models
wget -O results/pretrained_models/model_synth_only.mdl_pretrain https://files.icg.tugraz.at/f/e455628c1d8a44289a03/?dl=1

If you placed the model at a different location you need to change args.pretrained_model_filepath in config/config.py accordingly. You can then start the training from the pre-trained model by calling

cd source
python main_run.py --do-load-pretrained-model
Use fully trained models

Download the respective model from our server. We provide models trained with different numbers of labeled real samples. E.g., to use the model trained with all labeled real samples from the NYU dataset:

mkdir results/pretrained_models
wget -O results/pretrained_models/model_full_n73k.mdl https://files.icg.tugraz.at/f/71cd78dda19e4c4d8f67/?dl=1

You can then generate results using this one by:

cd source
python main_run.py --model-filepath ../results/pretrained_models/model_full_n73k.mdl --no-train

Training/Testing speed

In our case, data loading is the bottleneck. Hence, it's very beneficial if the data is stored on a disk with fast access times (e.g., SSD). Several workers are concurrently loading (and pre-processing) data samples. The number of workers can be changed by adjusting args.num_loader_workers in config/config.py.

Additionally, we use binary files to speed up training/testing. The binary files can be loaded faster, which will usually yield a significant speed up for training and testing.

Usage of the binary files can be toggled by setting args_data.use_pickled_cache (True/False) in config/config_data_nyu.py. If args_data.use_pickled_cache = True and a binary file for an image does not exist already, it is automatically written the first time the image is loaded. Hence, the process will be slower the first time training/testing is done with args_data.use_pickled_cache = True.

To ensure that all binary files will be properly written we recommend to generate all binary files before training/testing. To do so, call:

python main_generate_data_cache.py

Note, this is not always necessary but prevents possible issues during creation of the binary files.

Requirements

We used Python 2.7. To run the code you can, e.g., install the following requirements:

  • PyTorch (tested with version 0.4.1 and 1.0.1)
  • OpenCV (tested with version 2.4.9 and 3.1)
  • matplotlib
  • enum34
  • scipy
  • pycrayon
  • progressbar (only needed for main_generate_data_cache.py)

pycrayon

By default the code sends the data to port 8889 of "localhost". That is, you could start the server exactly as in the usage example in the crayon README, i.e., by calling

docker run -d -p 8888:8888 -p 8889:8889 --name crayon alband/crayon

See https://github.com/torrvision/crayon for details.

Citation

If you can make use of this work, please cite:

MURAUER: Mapping Unlabeled Real Data for Label AUstERity.
Georg Poier, Michael Opitz, David Schinagl and Horst Bischof.
In Proc. WACV, 2019.

Bibtex:

@inproceedings{Poier2019wacv_murauer,  
  author = {Georg Poier and Michael Opitz and David Schinagl and Horst Bischof},  
  title = {{MURAUER}: Mapping Unlabeled Real Data for Label AUstERity},  
  booktitle = {{Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV)}},  
  year = {2019}
}

References

[1] https://cims.nyu.edu/~tompson/NYU_Hand_Pose_Dataset.htm