Patchmentation Dataset

This datasets are used to benchmark patch augmentation performance of patchmentation.

The benchmarking results can be found at Xu-Justin/patchmentation-yolov5.

Dependency

• Using PIP

  pip install -r requirements.txt

• Using Docker (recommended)

  docker pull jstnxu/patchmentation:dataset
  docker run -it \
    -v {cache_folder}:/root/.cache/patchmentation-data \
    -v {data_folder}:/workspace/data \
    jstnxu/patchmentation:dataset /bin/bash

  • change {cache_folder} to local path to save cache.

  • change {data_folder} to local path to save generated data.

Dataset Spesification

• Training Dataset

  train-pascal-voc-2007

  • Number of Images: 2501

  • Number of Classes: 20

  • Source: Pascal VOC 2007 - Train

  python3 dataset.py --version train-pascal-voc-2007 --generate

  train-pascal-voc-2007-tiny

  • Number of Images: 200

  • Number of Classes: 20

  • Source: Pascal VOC 2007 - Train

  python3 dataset.py --version train-pascal-voc-2007-tiny --generate --batch 2

  train-pascal-voc-2007-v1

  • Number of Images: 2,500

  • Number of Classes: 20

  • Source: Pascal VOC 2007 - Train

  • Actions

    • filter.FilterWidth(50, Comparator.GreaterEqual)

    • filter.FilterHeight(50, Comparator.GreaterEqual)

    • transform.RandomResize(width_range=(50, 150), aspect_ratio=transform.Resize.AUTO_ASPECT_RATIO)

  • Kwargs

    • max_n_patches = 10

  
  

  python3 dataset.py --version train-pascal-voc-2007-v1 --generate --batch 30

  train-pascal-voc-2007-v2

  • Number of images: 2,500

  • Number of Classes: 20

  • Source: Pascal VOC 2007 - Train

  • Actions

    • filter.FilterWidth(50, Comparator.GreaterEqual)

    • filter.FilterHeight(50, Comparator.GreaterEqual)

    • transform.RandomResize(width_range=(50, 150), aspect_ratio=transform.Resize.AUTO_ASPECT_RATIO)

    • filter.FilterWidth(30, Comparator.GreaterEqual)

    • filter.FilterHeight(30, Comparator.GreaterEqual)

    • transform.SoftEdge(13, 20)

  • Kwargs

    • max_n_patches = 20

    • visibility_threshold = 1.0

  
  

  python3 dataset.py --version train-pascal-voc-2007-v2 --generate --batch 30

  train-pascal-voc-2007-v3

  • Number of images: 2,500

  • Number of Classes: 20

  • Source: Pascal VOC 2007 - Train

  • Actions

    • filter.FilterWidth(50, Comparator.GreaterEqual)

    • filter.FilterHeight(50, Comparator.GreaterEqual)

    • transform.RandomResize(width_range=(50, 150), aspect_ratio=transform.Resize.AUTO_ASPECT_RATIO)

  • Kwargs

    • max_n_patches = 20

    • visibility_threshold = 0.8

    • ratio_negative_patch = 5.0

    • iou_negative_patch = 0.2

  
  

  python3 dataset.py --version train-pascal-voc-2007-v3 --generate --batch 30

  train-pascal-voc-2007-v4

  • Number of images: 2,500

  • Number of Classes: 20

  • Source: Pascal VOC 2007 - Train

  • Actions

    • filter.FilterWidth(50, Comparator.GreaterEqual)

    • filter.FilterHeight(50, Comparator.GreaterEqual)

    • transform.RandomResize(width_range=(50, 150), aspect_ratio=transform.Resize.AUTO_ASPECT_RATIO)

    • filter.FilterWidth(30, Comparator.GreaterEqual)

    • filter.FilterHeight(30, Comparator.GreaterEqual)

    • transform.SoftEdge(13, 20)

  • Kwargs

    • max_n_patches = 20

    • visibility_threshold = 0.8

    • ratio_negative_patch = 5.0

    • iou_negative_patch = 0.2

  
  

  python3 dataset.py --version train-pascal-voc-2007-v4 --generate --batch 30

  train-penn-fudan-ped-person

  • Number of images: 100

  • Number of Classes: 1

  • Source: Penn Fudan Ped

  python3 dataset.py --version train-penn-fudan-ped-person --generate --batch 100

  train-campus

  • Number of images: 250

  • Number of Classes: 1

  • Source: Campus - Garden1, Penn Fudan Ped

  • Actions

    • filter.FilterWidth(20, Comparator.GreaterEqual)

    • filter.FilterHeight(20, Comparator.GreaterEqual)

    • transform.RandomResize(height_range=(150, 600), aspect_ratio=transform.Resize.AUTO_ASPECT_RATIO)

    • transform.SoftEdge(5, 10)

  • Kwargs

    • max_n_patches = 30

    • visibility_threshold = 0.8

  
  

  python3 dataset.py --version train-campus --generate --batch 50

• Validation Dataset

  valid-pascal-voc-2007

  • Number of Images: 2,510

  • Number of Classes: 20

  • Source: Pascal VOC 2007 - Val

  python3 dataset.py --version valid-pascal-voc-2007 --generate

  valid-penn-fudan-ped-person

  • Number of images: 70

  • Number of Classes: 1

  • Source: Penn Fudan Ped

  python3 dataset.py --version valid-penn-fudan-ped-person --generate

  valid-campus

  • Number of images: 256

  • Number of Classes: 1

  • Source: Campus - Garden1

  python3 dataset.py --version valid-campus --generate

• Test Dataset

  test-pascal-voc-2007

  • Number of Images: 4,952

  • Number of Classes: 20

  • Source: Pascal VOC 2007 - Test

  python3 dataset.py --version test-pascal-voc-2007 --generate

  test-campus

  • Number of images: 11,538

  • Number of Classes: 1

  • Source: Campus - Garden1

  python3 dataset.py --version test-campus --generate

Arguments

Priority*	Arguments	Type	Description
-	--version	one or more str	Dataset version(s).
-	--overwrite	store_true	Overwrite existing dataset / zip.
-	--batch	int	Number of batch to generate (default=1)
1	--generate	store_true	Generate the dataset. If overwrite is true, it will remove the dataset (if exists) before generating.
2	--zip	store_true	Zip the dataset. If overwrite is true, it will remove the dataset zip (if exists) before zipping.
3	--upload	store_true	Upload the dataset zip.
4	--remove-zip	store_true	Remove the dataset zip, if exists.
5	--download	one or more url	Download the dataset zip. If overwrite is true, it will remove the dataset zip (if exists) before downloading.
6	--unzip	store_true	Unzip the dataset zip. If overwrite is true, it will remove the dataset (if exists) before unzipping.
7	--validate	store_true	Validate the dataset.
8	--remove	store_true	Remove the dataset, if exists.

*Smaller priority number will be executed first

---

This project was developed as part of thesis project, Computer Science, BINUS University.