example.yaml

# Dataset name (case-sensitive). To create a new dataset, you will need to 
data_name: '<dataset-name>'
# A list of data splits. helpful when dataset has multiple splits, 
# e.g., 'train', 'test', 'val' for YFCC.
data_split: ['<data-split0>', ..., '<data-splitN>']
data_root: '<data-root-path>'
# The program will create a directory to dump all the intermediate results,
# including the keypoints, regional features, local features and augmented features.
dump_root: '<dump-root-path>'
# By default, all dataset images will be processed ([0, null]). To perform on only several images,
# for example, set it to [0, 10] for the first ten images.
truncate: [0, null]

# Pretrained model path for *.pb files, or checkpoint files.
pretrained:
  loc_model: 'pretrained/<model-name>'
  reg_model: 'pretrained/retrieval_model'

# Configuration of regional feature extraction.
reg_feat:
  # Whether to extract the feature.
  infer: true
  # Whether to overwite the feature records in dump files.
  overwrite: false
  # The max dimension of input images (images exceeded it will be down-sized).
  max_dim: 1024

# Configuration of local feature extraction.
loc_feat:
  # Whether to extract the feature.
  infer: true
  # Whether to overwite the feature records in dump files.
  overwrite: false
  # The maximum number of features/keypoints.
  n_feature: <#features>
  # Batch size of local features.
  batch_size: 512
  # Whether to extract dense local feature. Enable it for dense-contextdesc.
  dense_desc: false
  # Detection threshold for SIFT.
  peak_thld: 0.04
  # The max dimension of input images (images exceeded it will be down-sized).
  max_dim: 1280
  # Whether to ignore the orientation estimation of SIFT. Enable it when images are not imposed
  # large rotation change.
  upright: true
  # Whether to extract image patches from image pyramid so as to improve the scale invariance.
  # Enable it when images are imposed large scale change.
  scale_diff: false

# Configuration of augmented feature extraction.
aug_feat:
  # Whether to extract the feature.
  infer: true
  # the dimensionality of reginal feature.
  reg_feat_dim: 2048
  # Whether to overwite the feature records in dump files.
  overwrite: false
  # Whether to quantize the saved features to uint8.
  quantz: false

# Configuration of formatted file generation.
post_format:
  # Whether to perform the formatting.
  enable: true
  # The suffix of saved files.
  suffix: '.contextdesc10k_upright'