Skip to content

Latest commit

 

History

History
119 lines (87 loc) · 9.54 KB

datasets.md

File metadata and controls

119 lines (87 loc) · 9.54 KB
Raw

Dataset Preparation Guide

If you want to use prepared configs to run the Accuracy Checker tool and the Model Quantizer, you need to organize <DATASET_DIR> folder with validation datasets in a certain way. Instructions for preparing validation data are described in this document.

Each dataset description consists of the following sections:

  • instruction for downloading the dataset
  • structure of <DATASET_DIR> that matches the dataset definition in the existing global configuration file (dataset_definitions.yml)
  • examples of using and presenting the dataset in the global configuration file

More detailed information about using predefined configuration files you can find here.

ImageNet

How download dataset

To download images from ImageNet, you need to have an account and agree to the Terms of Access. Follow the steps below:

  1. Go to the ImageNet homepage
  2. If you have an account, click Login. Otherwise, click Signup in the right upper corner, provide your data, and wait for a confirmation email
  3. Log in after receiving the confirmation email and go to the Download tab
  4. Select Download Original Images
  5. You will be redirected to the Terms of Access page. If you agree to the Terms, continue by clicking Agree and Sign
  6. Click one of the links in the Download as one tar file section to select it
  7. Unpack archive

To download annotation files, you need to follow the steps below:

  • val.txt
    1. Download archive
    2. Unpack val.txt from the archive caffe_ilsvrc12.tar.gz
  • val15.txt
    1. Download annotation file
    2. Rename ILSVRC2017_val.txt to val15.txt

Files layout

To use this dataset with OMZ tools, make sure <DATASET_DIR> contains the following:

  • ILSVRC2012_img_val - directory containing the ILSVRC 2012 validation images
  • val.txt - annotation file used for ILSVRC 2012
  • val15.txt - annotation file used for ILSVRC 2015

Datasets in dataset_definitions.yml

  • imagenet_1000_classes used for evaluation models trained on ILSVRC 2012 dataset with 1000 classes. (model examples: alexnet, vgg16)
  • imagenet_1000_classes_2015 used for evaluation models trained on ILSVRC 2015 dataset with 1000 classes. (model examples: se-resnet-152, se-resnext-50)
  • imagenet_1001_classes used for evaluation models trained on ILSVRC 2012 dataset with 1001 classes (background label + original labels). (model examples: googlenet-v2-tf, resnet-50-tf)

Common Objects in Context (COCO)

How download dataset

To download COCO dataset, you need to follow the steps below:

  1. Download 2017 Val images and 2017 Train/Val annotations
  2. Unpack archives

Files layout

To use this dataset with OMZ tools, make sure <DATASET_DIR> contains the following:

  • val2017 - directory containing the COCO 2017 validation images
  • instances_val2017.json - annotation file which used for object detection and instance segmentation tasks
  • person_keypoints_val2017.json - annotation file which used for human pose estimation tasks

Datasets in dataset_definitions.yml

  • ms_coco_mask_rcnn used for evaluation models trained on COCO dataset for object detection and instance segmentation tasks. Background label + label map with 80 public available object categories are used. Annotations are saved in order of ascending image ID.
  • ms_coco_detection_91_classes used for evaluation models trained on COCO dataset for object detection tasks. Background label + label map with 80 public available object categories are used (original indexing to 91 categories is preserved. You can find more information about object categories labels here). Annotations are saved in order of ascending image ID. (model examples: faster_rcnn_resnet50_coco, ssd_resnet50_v1_fpn_coco)
  • ms_coco_detection_80_class_with_background used for evaluation models trained on COCO dataset for object detection tasks. Background label + label map with 80 public available object categories are used. Annotations are saved in order of ascending image ID. (model examples: faster-rcnn-resnet101-coco-sparse-60-0001, ssd-resnet34-1200-onnx)
  • ms_coco_detection_80_class_without_background used for evaluation models trained on COCO dataset for object detection tasks. Label map with 80 public available object categories is used. Annotations are saved in order of ascending image ID. (model examples: ctdet_coco_dlav0_384, yolo-v3-tf)
  • ms_coco_keypoints used for evaluation models trained on COCO dataset for human pose estimation tasks. Each annotation stores multiple keypoints for one image. (model examples: human-pose-estimation-0001)
  • ms_coco_single_keypoints used for evaluation models trained on COCO dataset for human pose estimation tasks. Each annotation stores single keypoints for image, so several annotation can be associated to one image. (model examples: single-human-pose-estimation-0001)

WIDER FACE

How download dataset

To download WIDER Face dataset, you need to follow the steps below:

  1. Go to the WIDER FACE website
  2. Go to the Download section
  3. Select WIDER Face Validation images and download them from Google Drive or Tencent Drive
  4. Select and download Face annotations
  5. Unpack archives

Files layout

To use this dataset with OMZ tools, make sure <DATASET_DIR> contains the following:

  • WIDER_val - directory containing images directory
    • images - directory containing the WIDER Face validation images
  • wider_face_split - directory with annotation file
    • wider_face_val_bbx_gt.txt - annotation file

Datasets in dataset_definitions.yml

  • wider used for evaluation models on WIDER Face dataset where the face is the first class. (model examples: mtcnn, retinaface-resnet50)
  • wider_without_bkgr used for evaluation models on WIDER Face dataset where the face is class zero. (model examples: mobilefacedet-v1-mxnet)

Visual Object Classes Challenge 2012 (VOC2012)

How download dataset

To download VOC2012 dataset, you need to follow the steps below:

  1. Go to the VOC2012 website
  2. Go to the Development Kit section
  3. Select Download the training/validation data and download archive
  4. Unpack archive

Files layout

To use this dataset with OMZ tools, make sure <DATASET_DIR> contains the following:

  • VOCdevkit/VOC2012 - directory containing annotations, images, segmentation masks and image sets files directories
    • Annotations - directory containing the VOC2012 annotation files
    • JPEGImages - directory containing the VOC2012 validation images
    • ImageSets - directory containing the VOC2012 text files specifying lists of images for different tasks
      • Main/val.txt - image sets file for detection tasks
      • Segmentation/val.txt - image sets file for segmentation tasks
    • SegmentationClass - directory containing the VOC2012 segmentation masks

Datasets in dataset_definitions.yml

  • VOC2012 used for evaluation models on VOC2012 dataset for object detection task. Background label + label map with 20 object categories are used. (model examples: mobilenet-ssd, ssd300)
  • VOC2012_without_background used for evaluation models on VOC2012 dataset for object detection tasks. Label map with 20 object categories is used.(model examples: yolo-v2-ava-0001, yolo-v2-tiny-ava-0001)
  • VOC2012_Segmentation used for evaluation models on VOC2012 dataset for segmentation tasks. Background label + label map with 20 object categories are used.(model examples: deeplabv3)