Scaled-YOLOv4: Scaling Cross Stage Partial Network

We show that the YOLOv4 object detection neural network based on the CSP approach, scales both up and down and is applicable to small and large networks while maintaining optimal speed and accuracy. We propose a network scaling approach that modifies not only the depth, width, resolution, but also structure of the network... (read more)

PDF Abstract

Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT BENCHMARK
Real-Time Object Detection COCO YOLOv4-CSP-P7 MAP 55.4 # 1
FPS 16 # 15
Real-Time Object Detection COCO YOLOv4-tiny 416 MAP 21.7 # 20
FPS 371 # 1
Real-Time Object Detection COCO YOLOv4-CSP CD53s 640 MAP 47.5 # 4
FPS 73 # 3
Object Detection COCO test-dev YOLOv4 (CD53) box AP 45.5 # 40
AP50 64.1 # 44
AP75 49.5 # 43
APS 27 # 41
APM 49 # 36
APL 56.7 # 41
Object Detection COCO test-dev YOLOv4-P7 (CSP-P7, single-scale) box AP 55.4 # 3
AP50 73.3 # 4
AP75 60.7 # 3
APS 38.1 # 2
APM 59.5 # 2
APL 67.4 # 4
Object Detection COCO test-dev YOLOv4-P6 (CSP-P6, single-scale) box AP 54.3 # 7
AP50 72.3 # 8
AP75 59.5 # 7
APS 36.6 # 6
APM 58.2 # 5
APL 65.5 # 11
Object Detection COCO test-dev YOLOv4-P7 (CSP-P7, multi-scale) box AP 55.8 # 1
AP50 73.2 # 5
AP75 61.2 # 1
APS 38.8 # 1
APM 60.1 # 1
APL 68.2 # 1
Object Detection COCO test-dev YOLOv4-P5 (CSP-P5, single-scale) box AP 51.4 # 15
AP50 69.9 # 16
AP75 56.3 # 18
APS 33.1 # 19
APM 55.4 # 15
APL 62.4 # 19

Methods used in the Paper