Developing an object detector solution with Azure Custom Vision .NET SDK
Download as PPTX, PDF0 likes300 views
The document outlines a presentation on building an object detector solution using Azure's Custom Vision .NET SDK, explaining key concepts such as object detection, customization, and project creation. It details the steps required for training an object detector with images, utilizing the Google Open Images dataset, and emphasizes the importance of diverse training examples for model accuracy. Additionally, it provides resources and links to open-source projects available on GitHub for practical implementation.
![OIDv4ToCV project
• Custom Vision requires normalized coordinates [0, 1] (x_min,
y_min, width, height).
• Google Open Images Dataset v4 is normalized (x_min, x_max,
y_min, y_max).
• OIDv4 Toolkit generates denormalized coordinates (pixels in
the image) (x_min, y_min, x_max, y_max)
• OIDv4ToCV “de-denormalize” the coordinates for its use in
Custom Vision.
https://github.com/icebeam7/OIDv4ToCV](https://image.slidesharecdn.com/developinganobjectdetectorsolutionwithazurecustomvision-200611090637/85/Developing-an-object-detector-solution-with-Azure-Custom-Vision-NET-SDK-18-320.jpg)
Developing an object detector solution with Azure Custom Vision .NET SDK
- 1. Building an object detector solution with Azure Custom Vision .NET SDK Luis Beltrán
- 2. Thanks to
- 3. Luis Beltrán • Researcher at Tomas Bata University in Zlín, Czech Republic. • Lecturer at Tecnológico Nacional de México in Celaya, Mexico. • Xamarin, Azure & AI @darkicebeam luis@luisbeltran.mx Slides availables in https://bit.ly/LuisAzureDay
- 4. AGENDA • What is object detection? • What is Custom Vision Service? • Building an object detector – Steps • Demo • Final thoughts
- 5. Object Detection Object detection is the process of identifying elements of interest in videos and pictures as well as locating them in the image/video, typically through coordenadas of a bounding box that contains it.
- 7. Custom Vision Service • Customize image recognition or object detection to fit your business needs. • An AI service and end-to- end platform for applying computer vision to your specific scenario. customvision.ai
- 8. Building an Object Detector solution Create a Project Choose a domain Upload images Localize the object of interest Tag it (Label it) Train the detector Evaluate, publish & export the model
- 9. Custom Vision .NET SDK
- 10. Creating a project Custom Vision is organized hierarchically. At the top level, there is a project, which represents the data and model for a specific task. An object detector is a model built with Custom Vision using training images. Object Detector = Project
- 12. Choosing a domain When you create an object detection project, you select a domain, which optimizes the detector to recognize elements from specific subsets : • Brand Logos • Products on Shelves • General Moreover, there are Compact domain versions, which are “light” object detectors optimized for systems with time or processing constraints.
- 14. Uploading and tagging the images In order to create a high-precision object detector, Custom Vision Service needs several training images. A training image is a picture that the Custom Vision uses to identify the relevant object with a tag and its position inside the image.
- 15. Google Open Images Dataset V4 Open Images is a dataset of ~9M images that have been annotated with image-level labels, object bounding boxes and visual relationships. The training set of V4 contains 14.6M bounding boxes for 600 object classes on 1.74M images, making it the largest existing dataset with object location annotations. https://storage.googleapis.com/openimages/web/factsfigures_v4.html
- 16. Google Open Images Dataset V4 The boxes have been largely manually drawn by professional annotators in order to ensure accuracy and consistency. The images are very diverse and often contain complex scenes with several objects (8.4 per image on average), encouraging structural image annotations, such as visual relationships. The dataset is annotated with image-level labels spanning thousands of classes. https://storage.googleapis.com/openimages/web/factsfigures_v4.html
- 18. OIDv4ToCV project • Custom Vision requires normalized coordinates [0, 1] (x_min, y_min, width, height). • Google Open Images Dataset v4 is normalized (x_min, x_max, y_min, y_max). • OIDv4 Toolkit generates denormalized coordinates (pixels in the image) (x_min, y_min, x_max, y_max) • OIDv4ToCV “de-denormalize” the coordinates for its use in Custom Vision. https://github.com/icebeam7/OIDv4ToCV
- 23. Training the object detector Every time the object detector is (re)trained, a new iteration of the model is created.
- 25. Evaluate and publish the model • After the model is trained, it can be quickly tested with another image. The evaluation uses the most recently trained iteration. • If the model provides accurate results, it can be published, which allows the object detector: • To be accessed via an HTTP endpoint (or through the SDK) for online object detection (prediction). • To be exported to a platform for offline object detection.
- 29. Deploy anywhere! Run your models wherever you need them and according to your unique scenario & requirements. Export your trained models to devices or to containers for low-latency scenarios.
- 33. Demo: Developing an object detector
- 34. Uploading training images Project information Adding tags to the model Open-source project available on GitHub: https://github.com/icebeam7/ObjectDetectionCustomVisionNet
- 35. Training the model Publishing the model Open-source project available on GitHub: https://github.com/icebeam7/ObjectDetectionCustomVisionNet
- 36. Evaluating the model with test images Open-source project available on GitHub: https://github.com/icebeam7/ObjectDetectionCustomVisionNet
- 37. Exporting the model to several platforms Open-source project available on GitHub: https://github.com/icebeam7/ObjectDetectionCustomVisionNet
- 38. Project status in the Custom Vision portal after the training model was created Open-source project available on GitHub: https://github.com/icebeam7/ObjectDetectionCustomVisionNet
- 39. Demo: Using the object detector in mobile apps with Xamarin
- 40. Open-source project available on GitHub: https://github.com/icebeam7/ObjectDetectionMobile
- 41. Open-source project available on GitHub: https://github.com/icebeam7/ObjectDetectionMobile
- 42. Optimization tips • The best way to generate a quality object detector is to add more varied tagged images (different backgrounds, angles, object size, groups of photos, and variants of types). • Use at least 30 images per tag.
- 43. Optimization tips • Include images that truly represent what your classifier will encounter in the real world. Thus, photos in context are better than photos of objects in front of neutral backgrounds. • Always train your model after you have added new images.
- 44. Pricing
- 45. Examples of Custom Vision scenarios Additional Scenarios • Classify user images that are submitted to a website • Identifying elements – object counting, animal identification and more. • Hazard detection/industrial safety – adding custom rules to videos for real-time experiences Product line category detection Agriculture & Farming Plant Health Detection Items Identification on a Retail Website
- 46. Call to Action - Resources Custom Vision Service https://azure.microsoft.com/en-us/services/cognitive-services/custom-vision-service/ Custom Vision Docs https://docs.microsoft.com/es-mx/azure/cognitive-services/custom-vision-service/home Quickstart: Create an object detection project with the Custom Vision SDK https://docs.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/quickstarts/object- detection Custom Vision and TensorFlow https://docs.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/export-model-python
- 47. Question Time
- 48. Thank you for your attention! Luis Beltrán Tomás Bata University in Zlín Tecnológico Nacional de México en Celaya luis@luisbeltran.mx luisbeltran.mx @darkicebeam About Me: https://about.me/luis-beltran LinkedIn: https://linkedin.com/in/luisantoniobeltran SlideShare: https://slideshare.net/icebeam GitHub: https://github.com/icebeam7 YouTube: https://youtube.com/user/darkicebeam
- 49. Thank You!!!
- 50. Thanks to