https://github.com/mop/bier

Cleaned up reference implementation of BIER: Boosting Independent Embeddings Robustly.
https://github.com/mop/bier

cnn computer-vision embeddings

Last synced: 4 days ago
JSON representation

Cleaned up reference implementation of BIER: Boosting Independent Embeddings Robustly.

• Host: GitHub
• URL: https://github.com/mop/bier
• Owner: mop
• License: gpl-3.0
• Created: 2018-02-08T14:30:17.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2018-02-21T08:03:35.000Z (over 6 years ago)
• Last Synced: 2024-03-13T06:32:05.911Z (4 months ago)
• Topics: cnn, computer-vision, embeddings
• Language: Python
• Homepage:
• Size: 69.3 KB
• Stars: 39
• Watchers: 4
• Forks: 15
• Open Issues: 3
• Metadata Files:
  • Readme: Readme.md
  • License: LICENSE

Lists

• awesome-gradient-boosting-papers - [Code
• awesome-gradient-boosting-papers - [Code

README

        
This project is a cleaned up version of our PAMI submission 

"Deep Metric Learning with BIER: Boosting Independent Embeddings Robustly" in tensorflow. It extends our

original ICCV version with an adversarial auxiliary loss during training, which improves results. 

you are planning to use this work, please cite one of the following papers:

    @inproceedings{opitz2017iccv,

        author={M. Opitz and G. Waltner and H. Possegger and H. Bischof},

        booktitle={{ICCV}},

        title={{BIER: Boosting Independent Embeddings Robustly}},

        year={2017}

    }

    @article{opitz2018arxiv,

        author={M. Opitz and G. Waltner and H. Possegger and H. Bischof},

        journal={arXiv:cs/1801.04815},

        title={{Deep Metric Learning with BIER: Boosting Independent Embeddings Robustly}},

        year={2018}

    }

To run the code, see ./run.sh and ./run_eval.sh.

The train-images file is a numpy file consisting of images of size 256x256 and train-labels are the corresponding

labels. The label indices should be between [0, total-number-of-labels) (i.e. they should be non-negative, and continuous).

The initial weights for the GoogLeNet InceptionV1 model (dumped from the caffe model) is available at: https://files.icg.tugraz.at/d/c3e1a047fc/