PP-YOLOE: An evolved version of YOLO

In this report, we present PP-YOLOE, an industrial state-of-the-art object detector with high performance and friendly deployment. We optimize on the basis of the previous PP-YOLOv2, using anchor-free paradigm, more powerful backbone and neck equipped with CSPRepResStage, ET-head and dynamic label assignment algorithm TAL. We provide s/m/l/x models for different practice scenarios. As a result, PP-YOLOE-l achieves 51.4 mAP on COCO test-dev and 78.1 FPS on Tesla V100, yielding a remarkable improvement of (+1.9 AP, +13.35% speed up) and (+1.3 AP, +24.96% speed up), compared to the previous state-of-the-art industrial models PP-YOLOv2 and YOLOX respectively. Further, PP-YOLOE inference speed achieves 149.2 FPS with TensorRT and FP16-precision. We also conduct extensive experiments to verify the effectiveness of our designs. Source code and pre-trained models are available at https://github.com/PaddlePaddle/PaddleDetection.

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Object Detection BDD100K val PP-YOLOE mAP@0.5 59.7 # 1
2D Object Detection BDD100K val PP-YOLOE mAP 35.6 # 2
Object Detection COCO test-dev PP-YOLOE-m(CSPRepResNet-m, 640x640, single-scale ) box mAP 48.9 # 96
AP50 66.5 # 65
AP75 53.0 # 51
APS 28.6 # 64
APM 52.9 # 38
APL 63.8 # 30
Object Detection COCO test-dev PP-YOLOE-s(CSPRepResNet-s, 640x640, single-scale ) box mAP 43.1 # 159
AP50 60.5 # 124
AP75 46.6 # 105
APS 23.2 # 113
APM 46.4 # 97
APL 56.9 # 82
Object Detection COCO test-dev PP-YOLOE-l(CSPRepResNet-l, 640x640, single-scale ) box mAP 51.4 # 75
AP50 68.9 # 45
AP75 55.6 # 38
APS 31.4 # 39
APM 55.3 # 25
APL 66.1 # 19
Object Detection COCO test-dev PP-YOLOE-x(CSPRepResNet-x, 640x640, single-scale ) box mAP 52.2 # 69
AP50 69.9 # 36
AP75 56.5 # 32
APS 33.3 # 28
APM 56.3 # 18
APL 66.4 # 17
Multiple Object Tracking CroHD PP-Tracking MOTA 72.6 # 1
Online Multi-Object Tracking MOT16 PP-Tracking MOTA 77.7 # 1
Multi-Object Tracking MOT16 PPTracking MOTA 77.7 # 1
Real-Time Object Detection MS COCO PP-YOLOE+_L FPS (V100, b=1) 78 # 13
box AP 52.9 # 20
FPS 78 # 11
Real-Time Object Detection MS COCO YOLOv3 FPS (V100, b=1) 123 # 6
box AP 51.0 # 26
FPS 123 # 5
Real-Time Object Detection MS COCO PP-YOLOE+_M FPS (V100, b=1) 123 # 6
box AP 49.8 # 29
FPS 123 # 5
Real-Time Object Detection MS COCO PP-YOLOE+_L(640,distillation) FPS (V100, b=1) 78 # 13
box AP 54.0 # 14
FPS 78 # 11
Real-Time Object Detection MS COCO PP-YOLOE+_X(640) FPS (V100, b=1) 45 # 18
box AP 54.7 # 11
FPS 45 # 18
Object Detection VisDrone-DET2019 PP-YOLOE-plus AP50 66.7 # 1

Methods