`ultralytics 8.0.149` add Open Images V7 training (#4178)
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: AdiEcho <30563671+AdiEcho@users.noreply.github.com>single_channel
Glenn Jocher
1 year ago
committed by
GitHub
parent
c751c7f88a
commit
7565210484
@ -0,0 +1,106 @@
|
|||||||
|
---
|
||||||
|
comments: true
|
||||||
|
description: Dive into Google's Open Images V7, a comprehensive dataset offering a broad scope for computer vision research. Understand its usage with deep learning models.
|
||||||
|
keywords: Open Images V7, object detection, segmentation masks, visual relationships, localized narratives, computer vision, deep learning, annotations, bounding boxes
|
||||||
|
---
|
||||||
|
|
||||||
|
# Open Images V7 Dataset
|
||||||
|
|
||||||
|
[Open Images V7](https://storage.googleapis.com/openimages/web/index.html) is a versatile and expansive dataset championed by Google. Aimed at propelling research in the realm of computer vision, it boasts a vast collection of images annotated with a plethora of data, including image-level labels, object bounding boxes, object segmentation masks, visual relationships, and localized narratives.
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
## Key Features
|
||||||
|
|
||||||
|
- Encompasses ~9M images annotated in various ways to suit multiple computer vision tasks.
|
||||||
|
- Houses a staggering 16M bounding boxes across 600 object classes in 1.9M images. These boxes are primarily hand-drawn by experts ensuring high precision.
|
||||||
|
- Visual relationship annotations totaling 3.3M are available, detailing 1,466 unique relationship triplets, object properties, and human activities.
|
||||||
|
- V5 introduced segmentation masks for 2.8M objects across 350 classes.
|
||||||
|
- V6 introduced 675k localized narratives that amalgamate voice, text, and mouse traces highlighting described objects.
|
||||||
|
- V7 introduced 66.4M point-level labels on 1.4M images, spanning 5,827 classes.
|
||||||
|
- Encompasses 61.4M image-level labels across a diverse set of 20,638 classes.
|
||||||
|
- Provides a unified platform for image classification, object detection, relationship detection, instance segmentation, and multimodal image descriptions.
|
||||||
|
|
||||||
|
## Dataset Structure
|
||||||
|
|
||||||
|
Open Images V7 is structured in multiple components catering to varied computer vision challenges:
|
||||||
|
|
||||||
|
- **Images**: About 9 million images, often showcasing intricate scenes with an average of 8.3 objects per image.
|
||||||
|
- **Bounding Boxes**: Over 16 million boxes that demarcate objects across 600 categories.
|
||||||
|
- **Segmentation Masks**: These detail the exact boundary of 2.8M objects across 350 classes.
|
||||||
|
- **Visual Relationships**: 3.3M annotations indicating object relationships, properties, and actions.
|
||||||
|
- **Localized Narratives**: 675k descriptions combining voice, text, and mouse traces.
|
||||||
|
- **Point-Level Labels**: 66.4M labels across 1.4M images, suitable for zero/few-shot semantic segmentation.
|
||||||
|
|
||||||
|
## Applications
|
||||||
|
|
||||||
|
Open Images V7 is a cornerstone for training and evaluating state-of-the-art models in various computer vision tasks. The dataset's broad scope and high-quality annotations make it indispensable for researchers and developers specializing in computer vision.
|
||||||
|
|
||||||
|
## Dataset YAML
|
||||||
|
|
||||||
|
Typically, datasets come with a YAML (Yet Another Markup Language) file that delineates the dataset's configuration. For the case of Open Images V7, a hypothetical `OpenImagesV7.yaml` might exist. For accurate paths and configurations, one should refer to the dataset's official repository or documentation.
|
||||||
|
|
||||||
|
!!! example "OpenImagesV7.yaml"
|
||||||
|
|
||||||
|
```yaml
|
||||||
|
--8<-- "ultralytics/cfg/datasets/open-images-v7.yaml"
|
||||||
|
```
|
||||||
|
|
||||||
|
## Usage
|
||||||
|
|
||||||
|
To train a YOLOv8n model on the Open Images V7 dataset for 100 epochs with an image size of 640, you can use the following code snippets. For a comprehensive list of available arguments, refer to the model [Training](../../modes/train.md) page.
|
||||||
|
|
||||||
|
!!! warning
|
||||||
|
|
||||||
|
The complete Open Images V7 dataset comprises 1,743,042 training images and 41,620 validation images, requiring approximately **561 GB of storage space** upon download.
|
||||||
|
|
||||||
|
Executing the commands provided below will trigger an automatic download of the full dataset if it's not already present locally. Before running the below example it's crucial to:
|
||||||
|
|
||||||
|
- Verify that your device has enough storage capacity.
|
||||||
|
- Ensure a robust and speedy internet connection.
|
||||||
|
|
||||||
|
!!! example "Train Example"
|
||||||
|
|
||||||
|
=== "Python"
|
||||||
|
|
||||||
|
```python
|
||||||
|
from ultralytics import YOLO
|
||||||
|
|
||||||
|
# Load a COCO-pretrained YOLOv8n model
|
||||||
|
model = YOLO('yolov8n.pt')
|
||||||
|
|
||||||
|
# Train the model on the Open Images V7 dataset
|
||||||
|
model.train(data='open-images-v7.yaml', epochs=100, imgsz=640)
|
||||||
|
```
|
||||||
|
|
||||||
|
=== "CLI"
|
||||||
|
|
||||||
|
```bash
|
||||||
|
# Train a COCO-pretrained YOLOv8n model on the Open Images V7 dataset
|
||||||
|
yolo detect train data=open-images-v7.yaml model=yolov8n.pt epochs=100 imgsz=640
|
||||||
|
```
|
||||||
|
|
||||||
|
## Sample Data and Annotations
|
||||||
|
|
||||||
|
Illustrations of the dataset help provide insights into its richness:
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
- **Open Images V7**: This image exemplifies the depth and detail of annotations available, including bounding boxes, relationships, and segmentation masks.
|
||||||
|
|
||||||
|
Researchers can gain invaluable insights into the array of computer vision challenges that the dataset addresses, from basic object detection to intricate relationship identification.
|
||||||
|
|
||||||
|
## Citations and Acknowledgments
|
||||||
|
|
||||||
|
For those employing Open Images V7 in their work, it's prudent to cite the relevant papers and acknowledge the creators:
|
||||||
|
|
||||||
|
```bibtex
|
||||||
|
@article{OpenImages,
|
||||||
|
author = {Alina Kuznetsova and Hassan Rom and Neil Alldrin and Jasper Uijlings and Ivan Krasin and Jordi Pont-Tuset and Shahab Kamali and Stefan Popov and Matteo Malloci and Alexander Kolesnikov and Tom Duerig and Vittorio Ferrari},
|
||||||
|
title = {The Open Images Dataset V4: Unified image classification, object detection, and visual relationship detection at scale},
|
||||||
|
year = {2020},
|
||||||
|
journal = {IJCV}
|
||||||
|
}
|
||||||
|
```
|
||||||
|
|
||||||
|
A heartfelt acknowledgment goes out to the Google AI team for creating and maintaining the Open Images V7 dataset. For a deep dive into the dataset and its offerings, navigate to the [official Open Images V7 website](https://storage.googleapis.com/openimages/web/index.html).
|
||||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue