v9.1.1
Menu
Paper breakdownObject DetectionPRO2024

YOLOv11

An Overview of the Key Architectural Enhancements — C3k2, SPPF, C2PSA and Multi-Task Heads

Rahima Khanam, Muhammad Hussain

8

Concepts

Paper Overview

YOLOv11 (Khanam & Hussain, University of Huddersfield, October 2024) is a different kind of paper from most breakdowns on this site. YOLO11 itself was not introduced in a paper at all — Ultralytics (Jocher & Qiu) unveiled it at the YOLO Vision 2024 conference and shipped it as code. This arXiv report is the first written architectural analysis of that release: it reads the model apart, names its new modules, and places it in the nine-year YOLO lineage.

The analysis centres on three components. C3k2, a cheaper Cross Stage Partial block, replaces YOLOv8's C2f in the backbone, neck and head. SPPF (Spatial Pyramid Pooling — Fast) is retained at the end of the backbone. And C2PSA, a Cross Stage Partial block with spatial self-attention, is new — placed directly after SPPF, it is the attention mechanism YOLOv8 never had.

Around that sits the familiar backbone → neck → head layout, and an expanded task list: one architecture family serves object detection, instance segmentation, image classification, pose estimation, oriented bounding boxes (OBB) and tracking, each in nano, small, medium, large and extra-large sizes. That small end of the ladder matters outside benchmarks too: it is detectors of this class that let a home camera tell a person from a passing shadow instead of just flagging changed pixels, a gap explained in how Wi-Fi security cameras work.

The headline numbers come from Ultralytics' published benchmark, which the paper reproduces as its Figure 2: YOLO11m reaches higher COCO mAP than YOLOv8m with 22% fewer parameters, and YOLO11x reaches roughly 54.5% mAP50−95^{50-95} at about 13 ms on a T4 GPU — the top of the latency/accuracy frontier among YOLO versions. Read it as a guided tour of a production detector rather than a new method: the value is in seeing why each block is shaped the way it is.

Explainer video
Preview · first 30s

Chapter Roadmap

Click any topic to jump in

1
Nine Years of YOLO

From the 2015 single-stage detector to YOLOv10's NMS-free training — the lineage YOLO11 inherits from.

distilled into
2
Backbone, Neck, Head

YOLO11 keeps YOLOv8's three-part layout and swaps in three new modules: C3k2, SPPF + C2PSA.

upgraded with

Three new building blocks

3
C3k2 Block

A cheaper Cross Stage Partial block that replaces C2f everywhere — with a c3k switch for deeper inner modules.

4
SPPF

Three chained 5x5 max-pools reproduce 5/9/13 pyramid pooling at the end of the backbone.

5
C2PSA Attention

Self-attention over the coarsest feature map, wrapped in a CSP split — the module YOLOv8 lacks.

feeding
6
CBS Layers & Detect Head

Conv-BatchNorm-SiLU refinement and per-scale Conv2D branches that emit boxes and class scores.

deployed as

Capabilities and evidence

7
One Architecture, Five Tasks

Detection, segmentation, pose, oriented boxes and classification — n through x for each.

8
Results

~54.5% COCO mAP for 11x at ~13 ms; 11m beats YOLOv8m with 22% fewer parameters.

Concept by concept

8 sections, in reading order.

PRO

Unlock the full breakdown

All 8 concepts — figures, key points, formulas and interactive demos — are part of PixelBank Premium.

  1. 01Nine Years of YOLO
  2. 02Backbone, Neck, Head
  3. 03C3k2 — A Cheaper Cross Stage Partial Block
  4. 04SPPF — Spatial Pyramid Pooling, Fast
  5. 05C2PSA — Spatial Attention at the Coarsest Scale
  6. 06CBS Layers and the Detect Head
  7. 07One Architecture, Five Tasks
  8. 08Results