Giters

THU-MIG / yolov10

YOLOv10: Real-Time End-to-End Object Detection

Home Page:https://arxiv.org/abs/2405.14458

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

同样的环境下测试，yolov5n/s比yolov10n/s inference速度都要快

yf0994 opened this issue · comments

commented

用同一份数据集分别训练yolov5n/s和yolov10n/s，用pt和导出onnx测试，发现yolov5的inference速度比yolov10的inference速度要块不少，下面分别是yolov5n和yolov10n的测试结果：
yolov10n:
image 1/8 /home/data/backups/source/yolo/yolov10/../testimg/CH120240508-003928-701.mp4_20240621_181520.494.jpg: 288x512 2 motorcycles, 17.7ms
image 2/8 /home/data/backups/source/yolo/yolov10/../testimg/CH120240508-003928-701.mp4_20240621_181526.094.jpg: 288x512 1 motorcycle, 17.0ms
image 3/8 /home/data/backups/source/yolo/yolov10/../testimg/CH120240508-005028-712.mp4_20240621_182045.836.jpg: 288x512 1 bus, 2 suvs, 16.9ms
image 4/8 /home/data/backups/source/yolo/yolov10/../testimg/CH120240508-005028-712.mp4_20240621_182047.787.jpg: 288x512 1 bus, 3 suvs, 16.9ms
image 5/8 /home/data/backups/source/yolo/yolov10/../testimg/CH120240508-005028-712.mp4_20240621_182051.964.jpg: 288x512 1 bus, 3 suvs, 16.9ms
image 6/8 /home/data/backups/source/yolo/yolov10/../testimg/CH120240508-005228-714.mp4_20240621_182124.833.jpg: 288x512 2 suvs, 2 sedens, 16.9ms
image 7/8 /home/data/backups/source/yolo/yolov10/../testimg/CH120240508-010529-727.mp4_20240621_182613.612.jpg: 288x512 1 motorcycle, 1 suv, 1 pedestrian, 1 mpv, 16.9ms
image 8/8 /home/data/backups/source/yolo/yolov10/../testimg/CH120240508-010529-727.mp4_20240621_182617.387.jpg: 288x512 2 motorcycles, 1 suv, 1 seden, 2 mpvs, 17.0ms
Speed: 1.3ms preprocess, 17.0ms inference, 0.7ms postprocess per image at shape (1, 3, 288, 512)

yolov5n:
image 1/8 /home/data/backups/source/yolo/testimg/CH120240508-003928-701.mp4_20240621_181520.494.jpg: 288x512 2 motorcycles, 9.8ms
image 2/8 /home/data/backups/source/yolo/testimg/CH120240508-003928-701.mp4_20240621_181526.094.jpg: 288x512 1 motorcycle, 12.4ms
image 3/8 /home/data/backups/source/yolo/testimg/CH120240508-005028-712.mp4_20240621_182045.836.jpg: 288x512 1 bus, 2 suvs, 9.0ms
image 4/8 /home/data/backups/source/yolo/testimg/CH120240508-005028-712.mp4_20240621_182047.787.jpg: 288x512 1 bus, 3 suvs, 8.1ms
image 5/8 /home/data/backups/source/yolo/testimg/CH120240508-005028-712.mp4_20240621_182051.964.jpg: 288x512 1 bus, 3 suvs, 9.0ms
image 6/8 /home/data/backups/source/yolo/testimg/CH120240508-005228-714.mp4_20240621_182124.833.jpg: 288x512 2 suvs, 2 sedens, 8.2ms
image 7/8 /home/data/backups/source/yolo/testimg/CH120240508-010529-727.mp4_20240621_182613.612.jpg: 288x512 1 motorcycle, 1 suv, 1 pedestrian, 1 mpv, 8.2ms
image 8/8 /home/data/backups/source/yolo/testimg/CH120240508-010529-727.mp4_20240621_182617.387.jpg: 288x512 1 motorcycle, 1 suv, 1 seden, 1 mpv, 8.3ms
Speed: 3.1ms pre-process, 9.1ms inference, 40.9ms NMS per image at shape (1, 3, 288, 512)

commented

是否在推理的时候没有关闭一对多的头？

commented

是否在推理的时候没有关闭一对多的头？

没有关闭