HOW local descriptors

This is the official Python/PyTorch implementation of the HOW local descriptors from our ECCV 2020 paper:

@InProceedings{TJ20,
  author      = "Giorgos Tolias and Tomas Jenicek and Ond\v{r}ej Chum}",
  title       = "Learning and aggregating deep local descriptors for instance-level recognition",
  booktitle   = "European Conference on Computer Vision",
  year        = "2020"
}

Running the Code

1. Install the cirtorch package (see cirtorch github for details)

# cirtorch
wget "https://github.com/filipradenovic/cnnimageretrieval-pytorch/archive/v1.2.zip"
unzip v1.2.zip
rm v1.2.zip
export PYTHONPATH=${PYTHONPATH}:$(realpath cnnimageretrieval-pytorch-1.2)

2. Install the asmk package with dependencies (see asmk github for details)

# asmk
git clone https://github.com/jenicek/asmk.git
pip3 install pyaml numpy faiss-gpu
cd asmk
python3 setup.py build_ext --inplace
rm -r build
cd ..
export PYTHONPATH=${PYTHONPATH}:$(realpath asmk)

3. Install pip3 requirements

pip3 install -r requirements.txt

4. Run examples/demo_how.py with two arguments – mode (train or eval) and any .yaml parameter file from examples/params/*/*.yml

Evaluating ECCV 2020 HOW models

Reproducing results from Table 2. with the publicly available models

• R18_how (n = 1000):   examples/demo_how.py eval examples/params/eccv20/eval_how_r18_1000.yml -e official_how_r18_1000   ROxf (M): 75.1, RPar (M): 79.4
• -R50_how (n = 1000):   examples/demo_how.py eval examples/params/eccv20/eval_how_r50-_1000.yml -e official_how_r50-_1000   ROxf (M): 78.3, RPar (M): 80.1
• -R50_how (n = 2000):   examples/demo_how.py eval examples/params/eccv20/eval_how_r50-_2000.yml -e official_how_r50-_2000   ROxf (M): 79.4, RPar (M): 81.6

Training HOW models

• R18_how:
  • train: examples/demo_how.py train examples/params/eccv20/train_how_r18.yml -e train_how_r18
  • eval (n = 1000): examples/demo_how.py eval examples/params/eccv20/eval_how_r18_1000.yml -ml train_how_r18
• -R50_how:
  • train: examples/demo_how.py train examples/params/eccv20/eval_how_r50-.yml -e train_how_r50-
  • eval (n = 1000): examples/demo_how.py eval examples/params/eccv20/eval_how_r50-_1000.yml -ml train_how_r50-
  • eval (n = 2000): examples/demo_how.py eval examples/params/eccv20/eval_how_r50-_2000.yml -ml train_how_r50-

Dataset shuffling during the training is done according to the cirtorch package; randomness in the results is caused by cudnn and by kmeans for codebook creation during evaluation.