Vision Datasets

Introduction

This repo

• defines unified contract for dataset for purposes such as training, visualization, and exploration, via DatasetManifest and ImageDataManifest.
• provides many commonly used dataset operation, such as sample dataset by categories, sample few-shot sub-dataset, sample dataset by ratios, train-test split, merge dataset, etc. (See here Link for available utilities)
• provides API for organizing and accessing datasets, via DatasetHub

Currently, six basic types of data are supported:

• classification_multiclass: each image can is only with one label.
• classification_multilabel: each image can is with one or multiple labels (e.g., 'cat', 'animal', 'pet').
• object_detection: each image is labeled with bounding boxes surrounding the objects of interest.
• image_caption: each image is labeled with a few texts describing the images.
• image_text_matching: each image is associated with a collection of texts describing the image, and whether each text description matches the image or not.
• image_matting: each image has a pixel-wise annotation, where each pixel is labeled as 'foreground' or 'background'.

multitask type is a composition type, where one set of images has multiple sets of annotations available for different tasks, where each task can be of any basic type.

Dataset Contracts

• DatasetManifest wraps the information about a dataset including labelmap, images (width, height, path to image), and annotations. ImageDataManifest encapsulates information about each image.
• ImageDataManifest encapsulates image-specific information, such as image id, path, labels, and width/height. One thing to note here is that the image path can be
  1. a local path (absolute c:\images\1.jpg or relative images\1.jpg)
  2. a local path in a non-compressed zip file (absolute c:\images.zip@1.jpg or relative images.zip@1.jpg) or
  3. an url
• ManifestDataset is an iterable dataset class that consumes the information from DatasetManifest.

ManifestDataset is able to load the data from all three kinds of paths. Both 1. and 2. are good for training, as they access data from local disk while the 3rd one is good for data exploration, if you have the data in azure storage.

For multitask dataset, the labels stored in the ImageDataManifest is a dict mapping from task name to that task's labels. The labelmap stored in DatasetManifest is also a dict mapping from task name to that task's labels.

Creating DatasetManifest

In addition to loading a serialized DatasetManifest for instantiation, this repo currently supports two formats of data that can instantiates DatasetManifest, using DatasetManifest.create_dataset_manifest(dataset_info, usage, container_sas_or_root_dir): COCO and IRIS (legacy).

DatasetInfo as the first arg in the arg list wraps the metainfo about the dataset like the name of the dataset, locations of the images, annotation files, etc. See examples in the sections below for different data formats.

Once a DatasetManifest is created, you can create a ManifestDataset for accessing the data in the dataset, especially the image data, for training, visualization, etc:

dataset = ManifestDataset(dataset_info, dataset_manifest, coordinates='relative')

Coco format

Here is an example with explanation of what a DatasetInfo looks like for coco format, when it is serialized into json:

    {
        "name": "sampled-ms-coco",
        "version": 1,
        "description": "A sampled ms-coco dataset.",
        "type": "object_detection",
        "format": "coco", // indicating the annotation data are stored in coco format
        "root_folder": "detection/coco2017_20200401", // a root folder for all files listed
        "train": {
            "index_path": "train.json", // coco json file for training, see next section for example
            "files_for_local_usage": [ // associated files including data such as images
                "images/train_images.zip"
            ]
        },
        "val": {
            "index_path": "val.json",
            "files_for_local_usage": [
                "images/val_images.zip"
            ]
        },
        "test": {
            "index_path": "test.json",
            "files_for_local_usage": [
                "images/test_images.zip"
            ]
        }
    }

Coco annotation format details w.r.t. multiclass/label_classification, object_detection, image_caption, image_text_match and multitask can be found in COCO_DATA_FORMAT.md.

Iris format

Iris format is a legacy format which can be found in IRIS_DATA_FORMAT.md. Only multiclass/label_classification, object_detection and multitask are supported.

Dataset management and access

Once you have multiple datasets, it is more convenient to have all the DatasetInfo in one place and instantiate DatasetManifest or even ManifestDataset by just using the dataset name, usage ( train, val ,test) and version.

This repo offers the class DatasetHub for this purpose. Once instantiated with a json including the DatasetInfo for all datasets, you can retrieve a ManifestDataset by

import pathlib

dataset_infos_json_path = 'datasets.json'
dataset_hub = DatasetHub(pathlib.Path(dataset_infos_json_path).read_text())
stanford_cars = dataset_hub.create_manifest_dataset(blob_container_sas, local_dir, 'stanford-cars', version=1, usage='train')

# note that you can pass multiple datasets.json to DatasetHub, it can combine them all
# example: DatasetHub([ds_json1, ds_json2, ...])
# note that you can specify multiple usages in create_manifest_dataset call
# example dataset_hub.create_manifest_dataset(blob_container_sas, local_dir, 'stanford-cars', version=1, usage=['train', 'val'])

for img, targets, sample_idx_str in stanford_cars:
    img.show()
    img.close()
    print(targets)

Note that this hub class works with data saved in both Azure Blob container and on local disk.

If local_dir:

1. is provided, the hub will look for the resources locally and download the data (files included in " files_for_local_usage", the index files, metadata (if iris format), labelmap (if iris format)) from blob_container_sas if not present locally
2. is NOT provided (i.e. None), the hub will create a manifest dataset that directly consumes data from the blob indicated by blob_container_sas. Note that this does not work, if data are stored in zipped files. You will have to unzip your data in the azure blob. (Index files requires no update, if image paths are for zip files: "a.zip@1.jpg"). This kind of azure-based dataset is good for large dataset exploration, but can be slow for training.

When data exists on local disk, blob_container_sas can be None.

Training with PyTorch

Training with PyTorch is easy. After instantiating a ManifestDataset, simply passing it in vision_datasets.pytorch.torch_dataset.TorchDataset together with the transform, then you are good to go with the PyTorch DataLoader for training.

Managing datasets with DatasetHub on cloud storage

If you are using DatasetHub to manage datasets in cloud storage, we recommend zipping (with uncompressed mode) the images into one or multiple zip files before uploading it and update the file path in index files to be like train.zip@train/1.jpg from train/1.jpg. For coco format, you can specify "file_name": "train/1.jpg" and "zip_file": "train.zip". You can do it with 7zip (set compression level to 'store') on Windows or zip command on Linux.

If you upload folders of images directly to cloud storage:

• you will need to list all images in "files_for_local_usage", which can be millions of entries
• downloading images one by one (even with multithreading) is much slower than downloading a few zip files