【ML Paper】YOLOv2: part8

This time, I'll introduce the YOLOv2 with the paper by Joseph Redmon and Ali Farhadi. Let's focus and see the difference from yolov1.

This article is part 8. Part 7 is here.

Original Paper: https://arxiv.org/abs/1612.08242

Faster

YOLOv2 is engineered for high-speed detection, crucial for applications such as robotics and self-driving cars that require low-latency predictions.

Traditional detection frameworks typically utilize VGG-16 as the base feature extractor, which, despite its robust classification capabilities, demands approximately $3.069 \times 10^{10}$ floating-point operations for processing a single image at a resolution of $224 \times 224$.

In contrast, YOLO employs a custom network inspired by the GoogLeNet architecture, reducing the computational load to $8.52 \times 10^{9}$ operations per forward pass.
This modification results in a slight decrease in accuracy, with YOLO achieving an 88.0% top-5 accuracy on ImageNet compared to VGG-16’s 90.0%, while significantly enhancing processing speed.

Darknet-19

To further optimize performance, YOLOv2 introduces Darknet-19 as its new classification model. Darknet-19 consists of 19 convolutional layers and 5 max-pooling layers, leveraging primarily $3 \times 3$ filters and doubling the number of channels after each pooling step, similar to VGG architectures.
Incorporating concepts from Network in Network (NIN), Darknet-19 utilizes global average pooling for predictions and integrates $1 \times 1$ filters to compress feature representations between $3 \times 3$ convolutions. Additionally, batch normalization is employed to stabilize training, accelerate convergence, and regularize the model.

This streamlined architecture requires only $5.58 \times 10^{9}$ floating-point operations to process an image, achieving 72.9% top-1 accuracy and 91.2% top-5 accuracy on the ImageNet dataset, thereby balancing computational efficiency with commendable classification performance.
For a full description see Table 6.