YOLO Model Task

YOLO Segmentation

Pixel-perfect instance segmentation that traces the exact outline of every detected object — the most detailed spatial data available from any real-time detection model.

4 min read Updated 25 May 2026

Article

What Is Instance Segmentation?

Instance segmentation is the computer vision task of identifying each object in an image and creating a pixel-perfect mask around it. While detection draws a rectangular bounding box, segmentation traces the exact outline — every pixel that belongs to the target is identified and separated from the background.

YOLO Segmentation combines real-time object detection with mask prediction in a single model pass. For each detected object, it outputs both a bounding box and a precise segmentation mask showing the exact silhouette.

The difference: A bounding box is a rectangle that inevitably includes background pixels. A segmentation mask is the exact shape of the target — providing the richest and most precise spatial information possible from a detection model.

What Segmentation Provides

Instance segmentation delivers capabilities that neither detection nor pose estimation alone can match:

Pixel-Perfect Boundaries

Know exactly which pixels on screen belong to a detected object and which don’t. The mask reveals the true shape rather than an approximate rectangle.

True Centre of Mass

Calculate the geometric centre of an object’s actual visible silhouette, not the centre of a rectangular bounding box that may include significant empty space.

Occlusion Awareness

When an object is partially hidden, the segmentation mask shows only the visible portion — revealing exactly how much of the target is actually exposed.

Visible Area Estimation

The mask’s pixel area reveals exactly how much of the target is visible on screen, providing data that bounding boxes simply cannot convey.

Detection vs Pose vs Segmentation

Each YOLO model task provides a different level of spatial information, forming a spectrum of detail:

Detection

Bounding boxes. Fastest inference. Location and size of each target in rectangular form.

Pose Estimation

Skeleton keypoints. Anatomical detail. Knows where specific body parts are located.

Segmentation

Pixel masks. Maximum spatial detail. Exact silhouette of each target at pixel level.

All Real-Time

All three tasks run in milliseconds on TensorRT-optimised GPUs. The choice depends on the scenario.

Trade-offs: Segmentation provides the richest spatial data but requires slightly more compute than detection alone. On modern RTX GPUs with TensorRT, this overhead is minimal — typically adding less than 1ms to inference time.

Gaming Applications

Segmentation is particularly powerful in these scenarios:

Cover & Partial Visibility

When a player is partially behind a wall, the segmentation mask shows only the exposed silhouette — far more useful than a bounding box that extends behind the cover.

Overlapping Targets

Instance segmentation identifies each player individually even when they overlap visually. Each gets their own distinct mask regardless of screen-space overlap.

Visual Clutter Filtering

The mask distinguishes target pixels from nearby particle effects, explosions, and ability visuals — separating signal from noise at the pixel level.

Stance Recognition

The shape of the segmentation mask reveals player stance — standing, crouching, prone, or mid-movement. The silhouette itself carries information about what the player is doing.

How YOLO Segmentation Works

At a high level, YOLO Segmentation extends the standard detection architecture:

  • Shared backbone — The same feature extraction network used for detection, meaning segmentation reuses most of the existing computation.
  • Detection head — Standard bounding box and class predictions, identical to detection mode.
  • Mask branch — An additional lightweight branch that generates segmentation masks for each detected instance.
  • Efficient design — Because the mask branch shares the backbone, the additional compute cost is minimal compared to running detection alone.

Responsible Use

NobleAIM is designed as an external computer vision tool for use in supported environments. We do not encourage or condone the use of our software in violation of any game’s Terms of Service. Please review our Terms of Service and Disclaimer for full details.

Pixel-Perfect Detection Technology

Experience the most precise spatial detection available, powered by YOLO Segmentation.