2. Network Architecture Explained. YOLO12 (YOLOv12) Series

Architectural flow in YOLOv12 (YOLO 12 series)

Input → Backbone (feature extraction)

• First two stages: regular convolution layers downsample the input and capture basic patterns (edges/textures).
• Repeated C3K2 modules: efficient feature refinement that builds hierarchical feature maps, capturing local + global patterns.
• Shallow vs deep behavior: shallow layers emphasize fine details, while deeper layers emphasize broader/high-level features.
• New A2C2F block: replaces/extends prior design elements from earlier YOLO versions.
  • Adds an area attention mechanism by splitting the feature map into sub-regions
  • Purpose: improve speed while keeping high detection accuracy
  • Focuses computation/attention on most relevant regions

Backbone → Neck (feature fusion)

• Takes multi-scale features from the backbone and refines them for detection.
• Uses upsampling to pull in finer details from earlier stages.
• Uses concat layers to combine fine + deep (abstract) features.
• A2C2F in the neck differs from backbone A2C2F:
  • In the neck, A2C2F blocks do not use the area attention mechanism
  • Instead, they behave more like C3K2-style feature merging
• Ends with one final C3K2 module completing the neck.

Neck → Head (prediction)

• The Detect module processes three output scales.
• Predicts:
  • Bounding boxes
  • Object classes
  • Confidence scores

Cross-model sharing and scaling details

• The described architecture is shared across all YOLO 12 model sizes (nano → extra-large).
• Tensor shapes scale based on configuration parameters:
  • MC: maximum number of channels
  • W: width multiplier
  • D: controls the number of loops inside C3K2 and A2C2F blocks
• The video emphasizes that a generic architecture can be used to compute tensor shapes for different sizes.

Example walkthrough (to understand the area attention blocks)

• The attention mechanism is explained using the YOLO 12 nano configuration:

  • D = 0.5
  • W = 0.25
  • MC = 1024

• Input example:

  • Original image size: 768 (height) × 1024 (width)
  • Letter boxing (aspect-ratio preserving resize + padding) to: 480 × 640
  • Network input becomes 480 × 640 × 3

• Notes on implementation detail:

  • Tensor dimensions are reordered for PyTorch-style format: from HWC to batch, channels, height, width (stated as “reorder dimensions from height width channels to batch channels height and width”).

YOLO 11 vs YOLO 12: key architectural changes

• Removed from YOLO 12:

  • SPF (SPatial Pyramid Pooling Fast layer)
  • C2PSA attention mechanism

• Replaced / redesigned attention backbone logic:

  • C3K2 is replaced with the new A2C2F area attention blocks
  • The video states A2C2F builds upon the prior attention idea (retains core attention flavor) but introduces the new area-based splitting strategy

• Overall message: YOLO 12 modifies several blocks/steps, prioritizing the area attention mechanism.

Main speakers / sources

• Primary speaker: the video presenter/author (no name provided in the subtitles)
• Referenced sources/models: Ultralytics YOLO 11 and Ultralytics YOLO 12 (YOLOv12 series)