ML Pipelines

Overview • How To Run

Overview

This application is a versatile tool designed for data transformation tasks (like filtering and augmentation). It allows you to create and manage Data transformation workflows by leveraging graphical nodes with settings.

Available Layers

Layers	Description	Images	Videos
Input
Images Project	Selects a project with images as the source data for processing through the pipeline.	+	-
Videos Project	Selects a project with videos as the source data for processing through the pipeline.	-	+
Input Labeling Job	Imports data and annotations from an existing labeling job as pipeline input.	+	-
Filtered Project	Uses a predefined project with pre-filtered images based on specific criteria.	+	-
Pixel Level Transformations
Anonymize	Applies pixelation or blurring effects to objects in images to hide sensitive information.	+	-
Blur	Applies various blur effects (Gaussian, median, etc.) to enhance or smooth image quality.	+	-
Contrast Brightness	Adjusts image contrast and brightness levels with precise control over parameters.	+	-
Noise	Adds controlled noise patterns to images for data augmentation or testing model robustness.	+	-
Random Color	Randomizes or systematically alters color values in images for augmentation purposes.	+	-
Spatial Level Transformations
Crop	Extracts specific regions from images based on configurable parameters or annotations.	+	-
Flip	Mirrors images horizontally or vertically while preserving annotation coordinates.	+	-
Instance Crop	Creates separate images for each detected object instance with configurable padding.	+	-
Multiply	Duplicates objects across the image with specified patterns and transformations.	+	-
Resize	Rescales images to target dimensions while properly transforming associated annotations.	+	-
Rotate	Rotates images by specified angles with proper transformation of associated annotations.	+	-
Sliding Window	Generates multiple overlapping crops from large images using a sliding window approach.	+	-
ImgAug Augmentations
ImgAug Studio	ImgAug Studio is a wrapper around ImgAug Library.	+	-
ImgAug.ImgCorruptlike.Noise	ImgAug imgcorruptlike Noise augmentators.	+	-
ImgAug.ImgCorruptlike.Blur	ImgAug imgcorruptlike Blur augmentators.	+	-
ImgAug.ImgCorruptlike.Weather	ImgAug imgcorruptlike Weather augmentators.	+	-
ImgAug.ImgCorruptlike.Color	ImgAug imgcorruptlike Color augmentators.	+	-
ImgAug.ImgCorruptlike.Compression	ImgAug imgcorruptlike Compression augmentators.	+	-
ImgAug.Geometric.ElasticTransformation	ImgAug geometric Elastic Transformation augmentator.	+	-
ImgAug.Geometric.PerspectiveTransform	ImgAug geometric Perspective Transform augmentator.	+	-
Annotation Transforms
Approximate Vector	Simplifies complex vector objects by reducing point count while preserving shape.	+	-
Background	Creates and assigns a background class to areas without object annotations.	+	+
Bounding Box	Converts any object annotation types to rectangular bounding boxes.	+	+
Bounding Box to Polygon	Converts rectangular bounding boxes to polygon annotations with configurable vertices.	+	+
Bitwise Masks	Performs logical operations (AND, OR, XOR) between masks of different classes.	+	-
Change Class Color	Modifies the display color of object classes without changing geometry or labels.	+	-
Drop Lines by Length	Removes line annotations based on their length using min/max thresholds.	+	-
Drop Noise	Filters out small mask fragments below specified area thresholds to reduce noise.	+	-
Drop Object by Class	Removes all objects of specified classes from annotations.	+	-
Duplicate Objects	Creates copies of selected objects with new class names while preserving geometry.	+	-
Image Tag	Adds custom metadata tags to images based on configurable conditions.	+	-
Line to Mask	Converts line annotations to mask annotations with configurable thickness.	+	-
Mask to Lines	Extracts the contours or center lines from mask annotations.	+	-
Mask to Polygon	Converts bitmap masks to polygon annotations with configurable precision.	+	-
Merge Classes	Combines multiple object classes into a single target class using an intuitive mapping table.	+	-
Merge Masks	Combines multiple mask annotations of the same class into a unified single mask.	+	-
Objects Filter	Selectively keeps or removes objects based on custom filtering criteria.	+	-
Objects Filter by Area	Tags or removes objects based on their area using configurable thresholds.	+	-
Polygon to Mask	Converts polygon annotations to bitmap mask representations.	+	-
Rasterize	Converts all vector annotations to bitmap mask representations.	+	-
Rename Classes	Renames object classes while preserving their original geometry and attributes.	+	-
Skeletonize	Creates skeletal representations (center lines) from mask annotations.	+	-
Split Masks	Separates connected mask regions into individual object instances.	+	-
Split Videos by Duration	Segments longer videos into smaller clips based on specified duration.	-	+
Filters and Conditions
Filter Image by Object	Includes or excludes images based on the presence of specific object classes.	+	-
Filter Image by Tag	Selects images that match specific tag criteria for further processing.	+	-
Filter Images without Objects	Identifies and processes only images that contain no object annotations.	+	-
Filter Videos by Duration	Selects videos based on their duration using min/max time thresholds.	-	+
Filter Videos by Objects	Includes or excludes videos based on the presence of specific object classes.	-	+
Filter Videos by Tags	Processes only videos that match specified tag criteria.	-	+
Filter Videos without Annotations	Identifies and processes videos that have no annotations.	-	+
Filter Videos without Objects	Selects videos that do not contain any object annotations.	-	+
IF Action	Creates conditional processing branches based on specified criteria or image properties.	+	-
Neural Networks
Apply NN Inference	Runs any deployed neural network model on images and incorporates results into the pipeline.	+	-
Deploy YOLOv5	Integrates and runs YOLOv5 object detection models with support for custom weights.	+	-
Deploy YOLO (v8, v9, v10, v11)	Deploys and runs multiple YOLO versions with support for detection, segmentation and classification tasks.	+	-
Deploy YOLO (v8 - v12)	Deploys the latest YOLO models with comprehensive support for all model variants and tasks.	+	-
Deploy MMDetection	Integrates the MMDetection library with access to numerous object detection architectures.	+	-
Deploy MMSegmentation	Runs semantic segmentation models from the MMSegmentation ecosystem with custom weights.	+	-
Deploy RT-DETR	Deploys Real-Time Detection Transformer models for efficient object detection.	+	-
Deploy RT-DETRv2	Integrates the improved version of RT-DETR with enhanced accuracy and performance.	+	-
Deploy DEIM	Deploys DEIM Models for efficient object detection.	+	-
Other
Dummy	Passes data through unchanged; useful for merging branches or as a placeholder.	+	-
Dataset	Consolidates all processed data into a single dataset with configurable naming.	+	-
Split Data	Distributes processed data across multiple datasets using various splitting strategies.	+	-
Output
Output Project	Saves processed data and annotations to a new or existing project with configurable settings.	+	-
Create new Project	Creates a new project to store the pipeline's output data and metadata.	+	+
Add to Existing Project	Appends processed data to a specified existing project with flexible dataset options.	+	+
Export Archive	Packages processed data and annotations as a downloadable archive in TeamFiles.	+	+
Export Archive with Masks	Exports data with separate mask files for each annotation in TeamFiles.	+	-
Copy Annotations	Transfers annotations between projects while maintaining original image references.	+	-
Create Labeling Job	Creates a new labeling job from processed data for further refinement.	+	+

Key features:

• Transform Data: Apply a wide variety of data transformation operations to images within a project. These transformations include rotation, cropping, blurring, resizing, and many more.

  

• Use Neural Networks: Apply deployed models on your data to perform object detection, instance segmentation, and other tasks. You can use any of the neural network models available in the Supervisely Ecosystem, or train your custom models.

  

• Enhance Data: Improve the quality and usability of your image data by adjusting contrast, brightness, and noise levels.

• Object-Level Manipulation: Perform operations on individual objects or instances within images, such as cropping, duplicating, or changing their color classes.

  

• Customize Workflows: Create complex data transformation workflows by combining multiple transformation nodes to meet your specific requirements.

• Node Documentation: Detailed documentation is available for each transformation node, explaining how to use it effectively. These guides provide step-by-step instructions and examples for each node, making it easy for users to understand and leverage the full power of the application.

  

• Save & Load Presets: Save your customized transformation workflows as presets for future use. This feature allows you to store and reuse your preferred configurations quickly.

  

• Output Flexibility: Choose from multiple export options to save your transformed data in a format that best suits your needs.

  

How To Run

There are several ways to run the application, depending on your needs and preferences:

1. Run App from Ecosystem

2. Run App from the context menu of the Project

3. Run App from the context menu of the Dataset

4. Run Pipeline from Project

5. Run Pipeline from Dataset

6. Run Pipeline with Filters