Skip to content

Latest commit

 

History

History
39 lines (33 loc) · 2.13 KB

DATASET.md

File metadata and controls

39 lines (33 loc) · 2.13 KB

Dataset

Type

The UG2+ 2022 Track2 provide 4 datasets:

  1. Normal-light Training Dataset (2625 videos w/ label)
  2. Normal-light Validation Dataset (330 videos w/ label)
  3. Dark Training Dataset (3088 videos w/o label)
  4. Dark Testing Dataset (3102 videos w/o label)

Besides, the organizers allow utilizing ARID dataset to validate and provide pseudo labels.

  1. ARID Training Dataset (6207 videos w/o label).

Prepartion

  1. Download all the data in UG2+ 2022 Track2 and ARID_v1.5.
  2. Prepare soft links in ./data, we haved provided all the CSV files.
# soft link for UG2
ln -s your_data_path/dark-train dark_train
ln -s your_data_path/dry-run dry_run
ln -s your_data_path/labeled-train labeled_train
ln -s your_data_path/Test Test
# copy data from ARID
ln -s your_data_path/ARID_v1.5/clips_v1.5 dark_train/Train

After the above steps, you can simply set DATA.PATH_PREFIX as data.

Usage

  1. [Supervised Training]:
    • Normal-light Training Dataset and Normal-light Validation Dataset are used for supervised training, all the videos and labels are utilized to train the models.
    • Dark Training Dataset and ARID Training Dataset are used for adapting BN, only the videos are utilized to update the parameters in BN.
    • ARID Training Dataset is used for validation, all the videos and labels are utilized to select the best model.
  2. [Semi-supervised Training]:
    • Dark Training Dataset and ARID Training Dataset are used for generating pseudo labels, only those pseudo labels with high confidence are utilized for training.
    • Normal-light Training Dataset and Normal-light Validation Dataset are also used for semi-supervised training.
    • ARID Training Dataset is still used for validation, all the videos and labels are utilized to select the best model.
  3. [Testing]
    • Dark Testing Dataset is used for testing, only the videos are used for generating corresponding predictions.