YOLO N Architecture Lecture

Introduction

• Overview of YOLO N architecture
• Explanation of key YOLO N blocks before diving into full architecture

Commonly Used Blocks

Convolutional Block

• Based on conv class in the common.py file (located in models folder)
• Consists of:
  • 2D convolutional layer
  • 2D batch normalization
  • Activation function (SiLU by default in YOLO 9)
• Employs auto padding calculation if not defined:
  • Formula: padding = kernel_size // 2
  • Example 1: kernel_size = 3, stride = 3, padding = 1
  • Example 2: kernel_size = 1, stride = 1, padding = 0

RepConv Block

• Based on RepConv class in common.py
• Components:
  • Two convolutional blocks
  • Element-wise addition
  • SiLU activation function
• Process:
  • Input passes through two convolutional blocks
  • First block: kernel_size = 3, padding = 1, stride = 1
  • Second block: kernel_size = 1, padding = 0, stride = 1
  • Element-wise addition of the two blocks' outputs
  • Apply SiLU activation function

RepBottleneck Block

• Based on RepBottleneck class in common.py
• Structure:
  • Sequence of blocks with shortcuts
  • Contains RepConv and convolutional blocks
• Similar to ResNet's bottleneck block (uses SiLU activation function)

RepCSP Block

• Based on RepCSP class in common.py
• Structure:
  • Three convolutional blocks
  • Sequence of RepBottleneck blocks
• Input split into two paths:
  • One to RepBottleneck
  • One to Concat block
• Ends with a convolutional block

RepCSPLayer Block

• Based on RepCSPLayer class in common.py
• Combines two architectures: CSPNet and ELAN
  • CSP: Cross Stage Partial Network
  • ELAN: Efficient Layer Aggregation Network
• Uses various computing blocks, not just convolutional layers
• Inputs split into two paths:
  • One through RepCSP and convolutional block
  • One directly to Concat block
• Ends with a convolutional block

SPPLayer Block

• Modification of SPP (Spatial Pyramid Pooling)
• Purpose: Generate feature representation of different object sizes without special info loss
• Structure:
  • Initial convolutional block
  • Three SPP blocks (each with max pool layer)
  • Concatenation before final convolutional block

YOLO N Architecture

General Structure

• Divided into three main parts: Backbone, Neck, and Head
• Introduces Programmable Gradient Information (PGI)
  • Provides complete input info for reliable gradient computation
  • Improves training process reliability
  • Added a new Auxiliary section

Detailed Breakdown

1. Backbone: Feature Extraction

   • Starts with silence block (no transformation)
   • Multiple convolutional blocks, RepCSPLayer blocks, and Adown blocks

2. Neck: Feature Combination

   • Upsample layer: Increases feature map resolution
   • Use Concat to combine feature maps from different blocks
   • Several RepCSPLayer and Adown blocks

3. Head: Prediction

   • Detect blocks specialized for small, medium, and large objects
   • Final detect block takes feature map from various preceding blocks

4. Auxiliary: Training Enhancement

   • Provides extra info linking input data to target output
   • Multiple blocks similar to Backbone for enhanced feature extraction
   • CBLinear and CBFuse blocks for higher-level features and feature fusion
   • Used only in training, can be deleted during inference

Conclusion

• Explanation of YOLO 9 architecture complete
• Further learning available through a three-in-one course covering YOLO 9, 8, and 7
• Links provided for additional resources

Thank you for watching!