Official implementation of "Sampling Wisely: Deep Image by Top-k Precision Optimization"(ICCV2019)

Citing this work

If you find this work useful in your research, please consider citing:

@inproceedings{luICCV19,
    Author = {Jing Lu and Chaofan Xu and Wei Zhang and Lingyu Duan and Tao Mei},
    Title = {Sampling Wisely: Deep Image by Top-k Precision Optimization},
    Booktitle = {IEEE International Conference on Computer Vision (ICCV)},
    Year = {2019}
}

DATASETS

• auto downloader and spliter
• CUB200-2011(with or without bounding box)
• CARS196(with or without bounding box)
• Stanford Online Product

LOSS

• smooth prec@k loss
• angular loss
• npair loss
• lifted loss
• semi-hard triplet loss
• contrastive loss
• arcface loss
• cosface loss(LMCL)
• proxyNCA
• A-softmax loss
• L-softmax loss
• softmax loss

VALIDATIONS

• precsion@k
• recall@k
• NMI
• F1
• mAP

TOOLS

• t-SNE visualization
• choose bad case

Installation

1. Install pytorch1.0, runconda install pytorch torchvision -c pytorch

2. Run conda install future requests six pillow

3. Run pip install sklearn tqdm

4. Run cd top_k_optimization

5. Choose right config file in main.py and set it whether you would download and split download and whether you need the dataset with bounding boxes in *_config.py and run python main.py

Validation

Draw t-SNE picture

Choose right config file in test_and_tsne.py and set it whether you would download and split download and whether you need the dataset with bounding boxes in *_config.py Run `python test_and_tsne.py'