Awesome video! Thanks a lot for exposing some important issues. To be honest, I am also tired of what is happening in the object detection space. It's nice to hear that I'm not crazy. ;)
This is true, everyone is stating that Yolo11 isn't a big deal. I'm performing many tests and will conclude how better it is. I have a RTX 3080 TI. There is only one reason why I'm working with YOLO11, it has extremely well documentation on how to get things running.
Hi, can you suggest which camera is best to buy for the Orange Pi 3B for video processing and object tracking? (Yolo+deepsort) I’ve always used the Logitech C270, but it’s getting old now. Is there anything better? As far as I remember, the Logitech also encoded in H.264.
I think it's obvious. Different GPU memory and bandwidth, architecture specific optimizations would be more optimized for some architectures and not the others. But the size of that difference is unknown until you try it.
AGX ORIN vs. RTX 4070 TI www.stereolabs.com/en-no/blog/performance-of-yolo-v5-v7-and-v8 (AGX ORIN vs RTX 4070) v5n: FPS(RTX 4070 TI)/FPS(AGX ORIN) = 2.5 v8n: FPS(RTX 4070 TI)/FPS(AGX ORIN) = 3 v5s: FPS(RTX 4070 TI)/FPS(AGX ORIN) = 3.1 Etc. My tests for different edge platforms - docs.google.com/spreadsheets/d/1BMj8WImysOSuiT-6O3g15gqHnYF-pUGUhi8VmhhAat4/edit?gid=0#gid=0 The difference for NPU can be dramatic. Here, at the end of this page, you can check the inference efficiency for different GPUs for Yolov5 lambdalabs.com/gpu-benchmarks So, as you can see - the difference in speed for different GPUs is usually ~20-30% For different NPUs, it can be 50-60% The main difference between GPUs and GPUs is technology. For example, in the last 2 generations, Nvidia using 3D tensor cores - ruclips.net/video/yyR0ZoCeBO8/видео.htmlsi=S1Ev_UZMAMZ5GUpk&t=20 And before, it was just an iteration. This resulted in a completely different speed different between different models
I've never though about models from this point view, very interesting, Thank you, video is excellent!
Glad it was helpful!
Awesome video! Thanks a lot for exposing some important issues. To be honest, I am also tired of what is happening in the object detection space. It's nice to hear that I'm not crazy. ;)
Thank you, Piotr!
This is true, everyone is stating that Yolo11 isn't a big deal. I'm performing many tests and will conclude how better it is. I have a RTX 3080 TI. There is only one reason why I'm working with YOLO11, it has extremely well documentation on how to get things running.
Hi, can you suggest which camera is best to buy for the Orange Pi 3B for video processing and object tracking? (Yolo+deepsort) I’ve always used the Logitech C270, but it’s getting old now. Is there anything better? As far as I remember, the Logitech also encoded in H.264.
Advise: If you say the new GPUs would make the data different, then, have few examples of it, otherwise, it is only an opinion. My 2 cents.
I think it's obvious. Different GPU memory and bandwidth, architecture specific optimizations would be more optimized for some architectures and not the others. But the size of that difference is unknown until you try it.
AGX ORIN vs. RTX 4070 TI
www.stereolabs.com/en-no/blog/performance-of-yolo-v5-v7-and-v8 (AGX ORIN vs RTX 4070)
v5n: FPS(RTX 4070 TI)/FPS(AGX ORIN) = 2.5
v8n: FPS(RTX 4070 TI)/FPS(AGX ORIN) = 3
v5s: FPS(RTX 4070 TI)/FPS(AGX ORIN) = 3.1
Etc.
My tests for different edge platforms - docs.google.com/spreadsheets/d/1BMj8WImysOSuiT-6O3g15gqHnYF-pUGUhi8VmhhAat4/edit?gid=0#gid=0
The difference for NPU can be dramatic.
Here, at the end of this page, you can check the inference efficiency for different GPUs for Yolov5
lambdalabs.com/gpu-benchmarks
So, as you can see - the difference in speed for different GPUs is usually ~20-30%
For different NPUs, it can be 50-60%
The main difference between GPUs and GPUs is technology. For example, in the last 2 generations, Nvidia using 3D tensor cores - ruclips.net/video/yyR0ZoCeBO8/видео.htmlsi=S1Ev_UZMAMZ5GUpk&t=20
And before, it was just an iteration. This resulted in a completely different speed different between different models