Latency comparison

[sunsmarterjie]

Thanks for the good work. Could you report the latency comparison with YOLO series method?

[yang-0201]

Thank you for your attention! During the model design process, I have considered optimizing speed; however, it is indeed challenging to balance accuracy, speed, and model size simultaneously.

We primarily focused on balancing accuracy and model size, so compared to other YOLO models, our model may be slower in terms of inference.

I have now conducted latency comparisons, using an L20 inference GPU and TensorRT 10.6.0 to test MHAF-YOLO, YOLO11, YOLOv10, and YOLOv9. The results are as follows:

Model	Speed L20 TensorRT10 (ms)	FPS	#Params (M)	FLOPs (G)	AP
MHAF-YOLO-N	1.28	930	2.2	7.2	42.3
YOLO11-N	1.01	1360	2.6	6.5	39.5
YOLOv10-N	0.92	1420	2.3	6.3	38.5
YOLOv9-T	1.53	773	2.0	7.7	38.3

Model	Speed L20 TensorRT10 (ms)	FPS	#Params (M)	FLOPs (G)	AP
MHAF-YOLO-S	1.67	686	7.1	25.3	48.9
YOLO11-S	1.01	1360	9.4	21.5	47.0
YOLOv10-S	1.23	979	7.2	21.6	46.3
YOLOv9-S	1.75	695	7.2	26.7	46.8

Model	Speed L20 TensorRT10 (ms)	FPS	#Params (M)	FLOPs (G)	AP
MHAF-YOLO-M	2.79	385	15.3	65.2	52.7
YOLO11-M	1.99	592	20.1	68.0	51.5
YOLOv10-M	1.80	624	15.4	59.1	51.1
YOLOv9-M	2.36	485	20.0	76.3	51.4

[sunsmarterjie]

Thanks

…

---- Replied Message ---- | From | Xiang ***@***.***> | | Date | 02/11/2025 01:27 | | To | ***@***.***> | | Cc | ***@***.***>***@***.***> | | Subject | Re: [yang-0201/MHAF-YOLO] Latency comparison (Issue #4) | Thank you for your attention! During the model design process, I have considered optimizing speed; however, it is indeed challenging to balance accuracy, speed, and model size simultaneously. We primarily focused on balancing accuracy and model size, so compared to other YOLO models, our model may be slower in terms of inference. I have now conducted latency comparisons, using an L20 inference GPU and TensorRT 10.6.0 to test MHAF-YOLO, YOLO11, YOLOv10, and YOLOv9. The results are as follows: | Model | Speed L20 TensorRT10 (ms) | FPS | #Params (M) | FLOPs (G) | AP | | MHAF-YOLO-N | 1.28 | 930 | 2.2 | 7.2 | 42.3 | | YOLO11-N | 1.01 | 1360 | 2.6 | 6.5 | 39.5 | | YOLOv10-N | 0.92 | 1420 | 2.3 | 6.3 | 38.5 | | YOLOv9-T | 1.53 | 773 | 2.0 | 7.7 | 38.3 | | Model | Speed L20 TensorRT10 (ms) | FPS | #Params (M) | FLOPs (G) | AP | | MHAF-YOLO-S | 1.67 | 686 | 7.1 | 25.3 | 48.9 | | YOLO11-S | 1.01 | 1360 | 9.4 | 21.5 | 47.0 | | YOLOv10-S | 1.23 | 979 | 7.2 | 21.6 | 46.3 | | YOLOv9-S | 1.75 | 695 | 7.2 | 26.7 | 46.8 | | Model | Speed L20 TensorRT10 (ms) | FPS | #Params (M) | FLOPs (G) | AP | | MHAF-YOLO-M | 2.79 | 385 | 15.3 | 65.2 | 52.7 | | YOLO11-M | 1.99 | 592 | 20.1 | 68.0 | 51.5 | | YOLOv10-M | 1.80 | 624 | 15.4 | 59.1 | 51.1 | | YOLOv9-M | 2.36 | 485 | 20.0 | 76.3 | 51.4 | — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>