Skip to content

Latest commit

 

History

History
329 lines (292 loc) · 9.35 KB

dataset_preparation.md

File metadata and controls

329 lines (292 loc) · 9.35 KB
Raw

How to prepare dataset for training

In this document, we will explain how to prepare dataset for training/inference for our purpose. We will assume you will gather these datset under ~/dataset directory. Note that you do not have to collect all dataset. If you would like to try our script as soon as possible, we recommend download RHD Dataset and Stereo Hand Pose Tracking Benchmark at least.

RGB Dataset

First-Person Hand Action Benchmark with RGB-D Videos and 3D Hand Pose Annotations (FHAD)

  1. Go to web site and press button Dataset download to get permission to download their dataset.
  2. Prepare dataset directory, for example, DATASET_DIR=~/dataset/fhad.
  3. Get Hand_pose_annotation_v1.zip and unzip in DATASET_DIR
  4. Get Video_files/Subject_X.zip (for X=1, 2, 3, 4, 5, 6) and unzip them in DATASET_DIR
  • By following the download procedure above, you'll get the following directory structure:
$ tree -d -L 2 ~/dataset/fhad
.
├── Hand_pose_annotation_v1
│   ├── Subject_1
│   ├── Subject_2
│   ├── Subject_3
│   ├── Subject_4
│   ├── Subject_5
│   └── Subject_6
└── Video_files
    ├── Subject_1
    ├── Subject_2
    ├── Subject_3
    ├── Subject_4
    ├── Subject_5
    └── Subject_6

Stereo Hand Pose Tracking Benchmark (a.k.a STB)

  1. Prepare dataset directory, for example, DATASET_DIR=~/dataset/stb
  2. Go to repository zhjwustc/icip17_stereo_hand_pose_dataset and click link of Dropbox i.e.
  • https://www.dropbox.com/sh/ve1yoar9fwrusz0/AAAfu7Fo4NqUB7Dn9AiN8pCca?dl=0
  1. Download all files e.g. BXCounting.zip, BXRandom.zip (for X = 1,2,3,4,5,6) or labels.zip and unzip them.
  • Note that since some folder doesn't generate directory, you'll need add -d option.
# for X = 1,2,3
$ unzip BXCounting.zip -d BXCounting
$ unzip BXRandom.zip -d BXRandom
# for X = 4,5,6
$ unzip BXCounting.zip
$ unzip BXRandom.zip
# finally unzip labels.zip
$ unzip labels.zip
  • By following the download procedure above, you'll get the following directory structure.
$ tree -d -L 2 ~/dataset/stb
.
├── images
│   ├── B1Counting
│   ├── B1Random
│   ├── B2Counting
│   ├── B2Random
│   ├── B3Counting
│   ├── B3Random
│   ├── B4Counting
│   ├── B4Random
│   ├── B5Counting
│   ├── B5Random
│   ├── B6Counting
│   └── B6Random
└── labels

GANerated Hands Dataset

  1. Prepare dataset directory, for example, ~/dataset
  2. Go to website https://handtracker.mpi-inf.mpg.de/projects/GANeratedHands/GANeratedDataset.htm and Download GANerated Hands Data: Compressed Zip: Single file (zip, 33.7 GB). You can get GANeratedDataset_v3.zip
  • Extract GANeratedDataset_v3.zip
$ cd ~/dataset
$ unzip GANeratedDataset_v3.zip
$ ls
GANeratedDataset_v3.zip GANeratedHands_Release
  • By following the download procedure above, you'll get the following directory structure.
$ tree -d -L 2 ~/dataset/GANeratedDataset_Release
.
└── data
    ├── noObject
    └── withObject

SynthHands

  1. Prepare dataset directory, for example, ~/dataset
  2. Go to website https://handtracker.mpi-inf.mpg.de/projects/OccludedHands/SynthHands.htm and Download Compressed Zip: Single file (zip, 48.5 GB). You can get SynthHands.zip.
  • Extract SynthHands.zip:
$ cd ~/dataset
$ unzip SynthHands.zip
$ ls
SynthHands.zip SynthHands_Release
  • By following the download procedure above, you'll get the following directory structure.
$ tree -d -L 2 ~/dataset/SynthHands_Release
├── female_noobject
│   ├── seq01
│   ├── seq02
│   ├── seq03
│   ├── seq04
│   ├── seq05
│   ├── seq06
│   └── seq07
├── female_object
│   ├── seq01
│   ├── seq02
│   ├── seq03
│   ├── seq04
│   ├── seq05
│   ├── seq06
│   └── seq07
├── male_noobject
│   ├── seq01
│   ├── seq02
│   ├── seq03
│   ├── seq04
│   ├── seq05
│   ├── seq06
│   └── seq07
└── male_object
    ├── seq01
    ├── seq02
    ├── seq03
    ├── seq04
    ├── seq05
    ├── seq06
    └── seq07

RHD dataset

Go to website and download dataset and unzip it.

$ unzip RHD_v1-1.zip
# you will see `RHD_published_v2` directory

You can see the following directory structure

$ tree -d -L 2 RHD_published_v2
RHD_published_v2/
├── evaluation
│   ├── color
│   ├── depth
│   └── mask
└── training
    ├── color
    ├── depth
    └── mask

Large-scale Multiview 3D Hand Pose Dataset

go to website and download dataset (see at the bottom of the website). You can get multiview_hand_pose_dataset_uploaded_v2.zip. Extract it and rename it as multiview_hand

You can confirm the directory structure as follow:

tree -d -L 2 multiview_hand
multiview_hand
├── annotated_frames
│   ├── data_1
│   ├── data_10
│   ├── data_11
│   ├── data_12
│   ├── data_13
│   ├── data_14
│   ├── data_15
│   ├── data_16
│   ├── data_17
│   ├── data_18
│   ├── data_19
│   ├── data_2
│   ├── data_20
│   ├── data_21
│   ├── data_3
│   ├── data_4
│   ├── data_5
│   ├── data_6
│   ├── data_7
│   ├── data_8
│   └── data_9
├── augmented_samples
│   ├── data_1
│   ├── data_11
│   ├── data_14
│   ├── data_15
│   ├── data_16
│   ├── data_17
│   ├── data_18
│   ├── data_19
│   ├── data_2
│   ├── data_21
│   ├── data_3
│   ├── data_4
│   ├── data_5
│   ├── data_6
│   ├── data_7
│   ├── data_8
│   └── data_9
├── calibrations
│   ├── data_1
│   ├── data_10
│   ├── data_11
│   ├── data_12
│   ├── data_13
│   ├── data_14
│   ├── data_15
│   ├── data_16
│   ├── data_17
│   ├── data_18
│   ├── data_19
│   ├── data_2
│   ├── data_20
│   ├── data_21
│   ├── data_3
│   ├── data_4
│   ├── data_5
│   ├── data_6
│   ├── data_7
│   ├── data_8
│   └── data_9
└── utils

Hand Keypoint Detection in Single Images using Multiview Bootstrapping

  1. Make directory named handdb_dataset and
  2. Go to website and download them at handdb_dataset:
  • Hands with Manual Keypoint Annotations
  • Hands from Synthetic Data (6546 + 3243 + 2348 + 2124 = 14261 annotations)
  • Hands from Panoptic Studio by Multiview Bootstrapping (14817 annotations)
  1. Move them at handdb_dataset and extract them.
  2. You will see the following structure:
$ tree -d -L 2 ~/dataset/handdb_dataset
handdb_dataset/
├── hand143_panopticdb
│   └── imgs
├── hand_labels
│   ├── manual_test
│   ├── manual_train
│   └── output_viz
└── hand_labels_synth
    ├── output_viz_synth
    ├── synth1
    ├── synth2
    ├── synth3
    └── synth4

FreiHAND Dataset

  1. Go to website and download dataset FreiHand Dataset.
  2. You will get FreiHAND_pub_v1.zip and extract it.
  3. You will see the following structure:
tree -L 2 -d FreiHAND_pub_v1
FreiHAND_pub_v1
├── evaluation
│   └── rgb
└── training
    ├── mask
    └── rgb

Depth Dataset

NYU Hand pose dataset

  1. Prepare dataset directory, for example, DATASET_DIR=~/dataset/nyu_hand_dataset_v2
  2. Go to website https://jonathantompson.github.io/NYU_Hand_Pose_Dataset.htm, download nyu_hand_dataset_v2.zip (92 GB)
  • Extract nyu_hand_dataset_v2.zip with the following command:
$ cd ~/dataset/nyu_hand_dataset_v2
$ unzip nyu_hand_dataset_v2.zip -d nyu_hand_dataset_v2
  • By following the download procedure above, you'll get the following directory structure.
$ tree -d -L 2 ~/dataset/nyu_hand_datset_v2
.
└── dataset
    ├── test
    └── train
  1. Finally, to specify this dataset for training or inference procedure, edit confing.ini to set nyu to type option of dataset section.

2015 MSRA Hand Gesture Dataset

  1. Go to website https://jimmysuen.github.io/ and download 2015 MSRA Hand Gesture Dataset.
  2. You will get cvpr15_MSRAHandGestureDB.zip and extract it.
  • By following the download procedure above, you'll get the following directory structure.
$ tree -d -L 1 ~/dataset/cvpr15_MSRAHandGestureDB
cvpr15_MSRAHandGestureDB
├── P0
├── P1
├── P2
├── P3
├── P4
├── P5
├── P6
├── P7
└── P8