Ultralytics YOLO27:
Get Started
Masks, keypoints, rotated boxes, and class labels — labeled, trained, turned into numbers, and shipped
Practitioner pathwayintermediate ~5 hours 10 lessons Final exam · Certificate

Beyond Detection: Segment, Pose, OBB, and Classify

You can train a detector. Ultralytics YOLO supports seven vision tasks, and this course teaches the five you will meet most often after detection — instance and semantic segmentation, pose, oriented bounding boxes, and classification — from label format and training to reading results in code, chaining models, and per-task evaluation and export. Depth estimation, the seventh task, is part of the task plan and the evaluation and export lessons.

By Ultralytics Academy

Begin courseSign in to save progress
What you'll learn

Choose the right Ultralytics YOLO26 task for a product question, then label, train, read, evaluate, and export segmentation, semantic, pose, OBB, and classification models, and chain them into pipelines that answer questions a detector cannot.

  • Map a product question to one of the seven Ultralytics YOLO tasks, its weights, its result field, and its gating metric.

  • Write and check polygon, PNG class-map, keypoint, four-corner OBB, and folder-layout datasets before training.

  • Train -seg, -sem, -pose, -obb, and -cls models and read mask mAP, mIoU, OKS-based pose mAP, rotated-box mAP, and top-1 accuracy.

  • Turn predictions into decisions: mask areas and overlaps, class-map coverage, joint angles, ship headings, and confidence-routed labels.

  • Chain tasks (detect then classify, pose plus tracking), keep their cost flat, and export every task to ONNX with a verified prediction.

What you'll build

  • A task plan that names the task, weights, label format, result field, and gating metric for each model in a pipeline.

  • Label checkers for segmentation and pose datasets that catch malformed rows before a GPU hour is spent.

  • One-epoch training runs for segmentation, semantic segmentation, pose, OBB, and classification, with each metric explained.

  • Scripts that measure mask area and overlap, walkable surface, elbow and knee angles, and ship axes from real predictions.

  • A two-stage pipeline that classifies each tracked person once, and ONNX exports of every task, each verified with a prediction.

Prerequisites

  • Complete Train your first YOLO model — this course assumes you can train, validate, and predict with a detector.

  • Comfortable with Python and NumPy arrays.

  • A GPU helps but is not required: every example runs on a laptop CPU at epochs=1 and a small imgsz.

Course content

4 modules · 10 lessons

Module 1

Plan
1

Pick the Task and Plan the Pipeline

Map each product question to one of the seven Ultralytics YOLO tasks, with its weights, result field, label format, and gating metric, and prove the plan by running every task on one image.

Module 2

Segment
2

Label and Train Instance Segmentation

Write polygon labels, check them in code, train yolo26n-seg.pt on COCO8-seg, and diagnose a model from the gap between its box and mask metrics.

3

Use Masks in Code: Area, Overlap, Outlines, Cutouts

Read masks.data and masks.xy correctly, measure object area and fill, test overlap with masks instead of boxes, and cut objects out of the frame.

4

Semantic Segmentation: Label Every Pixel

Predict a per-pixel class map with yolo26n-sem.pt, turn it into coverage numbers, build PNG class-map datasets, and read mIoU without being fooled by pixel accuracy.

Module 3

Pose and OBB
5

Label and Train Pose Models

Write keypoint labels with visibility, set kpt_shape and flip_idx correctly, train on COCO8-pose, and turn keypoints into confidence-gated joint angles.

6

Custom Keypoints and OKS Metrics

Design a keypoint schema for non-human objects, write and check its dataset YAML, and set OKS sigmas from how precisely each point can be placed.

7

Oriented Bounding Boxes for Rotated Objects

Label four-corner OBBs, train yolo26n-obb.pt on DOTA8, and turn xywhr predictions into a stable orientation axis without falling into the angle conventions.

Module 4

Classify, Combine, Ship
8

Train an Image Classifier

Lay out a folder dataset, train yolo26n-cls.pt on MNIST160, read probs and top-1 accuracy, and route low-confidence predictions to human review.

9

Combine Tasks in One Pipeline

Build a detect-then-classify cascade, classify each tracked object once instead of every frame, and turn pose keypoints into per-person time series.

10

Evaluate and Ship Every Task

Gate each task on its own metric at the deployment size, validate every task in one loop, and export every task to ONNX with a verified prediction.