Segmentation
Segmentation is the process of providing pixel-level precision for every object or region in an image, not just a rectangular bounding box. This is the most detailed type of visual annotation, producing a mask that outlines the exact shape of the object.
Segmentation is required when the boundaries between objects, or between an object and the background, are critically important.
Ideal for: Autonomous driving, medical imaging, robotics, scene understanding, geospatial analysis, and any scenario requiring precise object boundaries and pixel-perfect shape measurement.
¶ When to Use
Use Segmentation when you need pixel-level precision for every object or region in an image, not just a rectangular bounding box.
¶ Instance Segmentation Tasks (Object and Location)
Instance Segmentation separates individual, overlapping instances of objects and provides a distinct, pixel-accurate mask for each one.
| Input Type | Question Example | Purpose / Requirement |
|---|---|---|
| Image | “Outline the exact pixel boundaries of every vehicle and every person in the scene, assigning a unique ID to each instance.” | Autonomous Driving/Robotics: Accurate obstacle avoidance and path planning where precise object boundaries are essential. |
| Image | “Identify and trace the outline of every product bottle on the shelf, even when they overlap.” | Retail Automation: Accurate item counting and inventory check in crowded or complex visual environments. |
| Image | “Trace the precise boundary of each animal instance in the wildlife photo.” | Ecology/Conservation: Counting and tracking individual animals for population monitoring. |
¶ Semantic Segmentation Tasks (Region Understanding)
Semantic Segmentation assigns a class label to every single pixel in the image, grouping all pixels belonging to the same class into one region, regardless of individual instances.
| Input Type | Question Example | Purpose / Requirement |
|---|---|---|
| Image | “Label every pixel in the image as either ‘road,’ ‘sidewalk,’ ‘vegetation,’ or ‘building.’” | Scene Understanding: Creating a complete map of the environment for navigation systems. |
| Medical Scan | “Outline the precise boundary of the liver, kidneys, and spleen within the CT scan.” | Medical Imaging Analysis: Measuring the volume, size, and shape of organs or pathological regions (tumors). |
| Aerial Image | “Label all pixels belonging to ‘water,’ ‘forest,’ or ‘agricultural land.’” | Geospatial Analysis: Detailed land cover mapping and resource management. |
¶ Key Difference from Object Detection
| Feature | Object Detection | Segmentation |
|---|---|---|
| Localization | Uses a Bounding Box (rectangle). | Uses a Pixel-level Mask (polygons or contours). |
| Accuracy | Good for localization and counting. | Required for pixel-perfect shape and area measurement. |
| Output | A set of coordinates and class labels. | A detailed mask for the shape of the object. |
¶ When to Use Other Work Types
- Need simple bounding boxes? → Object Detection
- Need simple categorization? → Classification
- Need temporal tracking? → Event Tagging
¶ Supported Data Types
- Image - Pixel-level annotation of static images
¶ Getting Started
- Define your segmentation classes and requirements (instance vs semantic)
- Specify mask precision and overlap handling rules
- Prepare your image data
- Submit batches via API
- Track progress and download results with pixel masks
¶ How to Submit Batches
Indirect Demand (CSV File) - Best for large-scale projects with thousands of images.
Direct Demand (Inline Data) - Best for dynamic tasks or real-time processing.
¶ Best Practices
- Clear Class Definitions - Define precise criteria for each segmentation class
- Boundary Precision - Specify required accuracy for object boundaries
- Overlapping Regions - Provide clear guidelines for handling overlapping objects
- Instance vs Semantic - Choose the right segmentation type for your use case
- Quality Control - Review sample masks to ensure pixel-level accuracy and consistency