YOLOv4: Optimal Speed and Accuracy of Object Detection

23 Apr 2020Alexey BochkovskiyChien-Yao WangHong-Yuan Mark Liao

There are a huge number of features which are said to improve Convolutional Neural Network (CNN) accuracy. Practical testing of combinations of such features on large datasets, and theoretical justification of the result, is required... (read more)

PDF Abstract

Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT LEADERBOARD
Real-Time Object Detection COCO YOLOv4-608 MAP 43.5 # 5
FPS 62 # 2
Real-Time Object Detection COCO YOLOv4-512 MAP 43 # 6
FPS 83 # 1
Object Detection COCO test-dev YOLOv4-608 box AP 43.5 # 33
AP50 65.7 # 23
AP75 47.3 # 35
APS 26.7 # 32
APM 46.7 # 36
APL 53.3 # 46

Methods used in the Paper


METHOD TYPE
Mish
Activation Functions
DropBlock
Regularization
CSPResNeXt
Convolutional Neural Networks
CSPResNeXt Block
Skip Connection Blocks
Leaky ReLU
Activation Functions
Batch Normalization
Normalization
1x1 Convolution
Convolutions
Convolution
Convolutions
Global Average Pooling
Pooling Operations
Softmax
Output Functions
CSPDarknet53
Convolutional Neural Networks
Darknet-53
Convolutional Neural Networks
EfficientNet
Image Models
Inverted Residual Block
Skip Connection Blocks
Dense Connections
Feedforward Networks
Average Pooling
Pooling Operations
Dropout
Regularization
Swish
Activation Functions
Spatial Pyramid Pooling
Pooling Operations
Adaptive Feature Pooling
Pooling Operations
RoIAlign
RoI Feature Extractors
RPN
Region Proposal
PAFPN
Feature Extractors
Bottom-up Path Augmentation
Feature Extractors
FPN
Feature Extractors
Pointwise Convolution
Convolutions
Concatenated Skip Connection
Skip Connections
YOLOv3
Object Detection Models
k-Means Clustering
Clustering
Logistic Regression
Generalized Linear Models
YOLOv4
Object Detection Models
Polynomial Rate Decay
Learning Rate Schedules
Linear Warmup
Learning Rate Schedules
Weight Decay
Regularization
SGD with Momentum
Stochastic Optimization
Mixup
Image Data Augmentation
BiFPN
Feature Extractors
Depthwise Separable Convolution
Convolutions
ReLU
Activation Functions
RFB
Feature Extractors
Dilated Convolution
Convolutions
Residual Connection
Skip Connections
DIoU-NMS
Proposal Filtering
Label Smoothing
Regularization
Depthwise Convolution
Convolutions
CutMix
Image Data Augmentation
Sigmoid Activation
Activation Functions
Cosine Annealing
Learning Rate Schedules
Spatial Attention Module
Image Model Blocks
Max Pooling
Pooling Operations
ResNeXt Block
Skip Connection Blocks
Grouped Convolution
Convolutions