XuShangliang ; WangXinxin ; LvWenyu ; ChangQinyao ; CuiCheng ; DengKaipeng ; WangGuanzhong ; DangQingqing ; WeiShengyu ; DuYuning ; LaiBaohua

摘要
在本报告中,我们介绍了PP-YOLOE,这是一种具有高性能和友好部署的工业级最先进的目标检测器。我们在先前的PP-YOLOv2基础上进行了优化,采用了无锚点(anchor-free)范式,配备了更强的主干网络和颈部结构,其中包括CSPRepResStage、ET-head以及动态标签分配算法TAL。我们为不同的应用场景提供了s/m/l/x四种模型。结果表明,PP-YOLOE-l在COCO测试开发集上达到了51.4 mAP,在Tesla V100上实现了78.1 FPS的速度,相比之前的工业级最先进模型PP-YOLOv2和YOLOX分别提升了+1.9 AP(精度)和+13.35%的速度,以及+1.3 AP(精度)和+24.96%的速度。此外,使用TensorRT和FP16精度进行推理时,PP-YOLOE的推理速度达到了149.2 FPS。我们还进行了广泛的实验以验证设计的有效性。源代码和预训练模型可在以下链接获取:https://github.com/PaddlePaddle/PaddleDetection。
代码仓库
PaddlePaddle/PaddleYOLO
paddle
Nioolek/PPYOLOE_pytorch
pytorch
GitHub 中提及
Mind23-2/MindCode-130
mindspore
open-mmlab/mmyolo
pytorch
PaddlePaddle/PaddleDetection
官方
paddle
GitHub 中提及
Gaurav14cs17/YOLOE
pytorch
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| 2d-object-detection-on-1 | - | : |
| multi-object-tracking-on-mot16 | PPTracking | MOTA: 77.7 |
| multiple-object-tracking-on-crohd | PP-Tracking | MOTA: 72.6 |
| object-detection-on-4 | - | : |
| object-detection-on-coco | PP-YOLOE-l(CSPRepResNet-l, 640x640, single-scale ) | AP50: 68.9 AP75: 55.6 APL: 66.1 APM: 55.3 APS: 31.4 box mAP: 51.4 |
| object-detection-on-coco | PP-YOLOE-x(CSPRepResNet-x, 640x640, single-scale ) | AP50: 69.9 AP75: 56.5 APL: 66.4 APM: 56.3 APS: 33.3 box mAP: 52.2 |
| object-detection-on-coco | PP-YOLOE-s(CSPRepResNet-s, 640x640, single-scale ) | AP50: 60.5 AP75: 46.6 APL: 56.9 APM: 46.4 APS: 23.2 box mAP: 43.1 |
| object-detection-on-coco | PP-YOLOE-m(CSPRepResNet-m, 640x640, single-scale ) | AP50: 66.5 AP75: 53.0 APL: 63.8 APM: 52.9 APS: 28.6 box mAP: 48.9 |
| object-detection-on-visdrone-det2019-1 | PP-YOLOE-plus | AP50: 66.7 |
| online-multi-object-tracking-on-mot16 | PP-Tracking | MOTA: 77.7 |
| real-time-object-detection-on-coco | PP-YOLOE+_L(distillation) | FPS (V100, b=1): 78 box AP: 54.0 |
| real-time-object-detection-on-coco | PP-YOLOE+_M | box AP: 49.8 |
| real-time-object-detection-on-coco | PP-YOLOE+_L | FPS (V100, b=1): 78 box AP: 52.9 |
| real-time-object-detection-on-coco | PP-YOLOE+_X | FPS (V100, b=1): 45 box AP: 54.7 |
| real-time-object-detection-on-coco | YOLOv3 | FPS (V100, b=1): 123 box AP: 51.0 |