First, the confidence levels being different is simply a result of the different loss functions used. Modern yolos need much lower confidence values or longer training but as a result dont have as many false positives and false negatives and better bounding boxes. Second, the speed difference can be explained by the different image sizes used. 640x480 for Yolov10 are roughly 10 times more pixels than 224x160 for yolov3 and yolov4.
Why is there such a difference in speed? Is it the output on the gpu in both cases? What are the input sizes of the neural networks? And how much are the weights for each of them? I would be grateful for the answer.
I can only comment on Darknet/YOLO. I'm not familiar enough with Ultralytics to say why it is slower, other than to guess that Python is always slower than C and C++.
I didn't train this network to find people and crowds. It was trained to find the things you see in the video. Like all customers who ask me to train neural networks for them, they typically want to find specific objects in machinery, or on a conveyor belt, not MSCOCO-style "find 80 random classes of things."
@@StephaneCharette Okay, I watched the video, but shouldn't the new versions be better to yolov4 logically? (I use Google translate, sorry for any translation errors. )
@@g-4660-i4l The version bumps are just marketing. while there are technical changes, they do not provably provide obvious speed, accuracy or training improvements.
First, the confidence levels being different is simply a result of the different loss functions used. Modern yolos need much lower confidence values or longer training but as a result dont have as many false positives and false negatives and better bounding boxes.
Second, the speed difference can be explained by the different image sizes used. 640x480 for Yolov10 are roughly 10 times more pixels than 224x160 for yolov3 and yolov4.
Hey Stephane. Thanks for your time putting up this video. Results are certainly interesting. Keep up the great work that you do with darknet !
Thanks for your ongoing contribution to the fastest and most reliable object detection framework, Stephane!
hey nice video just wanna say thank u for the help with your channel life is easier
Another great video. Thanks Stephane!
Why is there such a difference in speed? Is it the output on the gpu in both cases? What are the input sizes of the neural networks? And how much are the weights for each of them? I would be grateful for the answer.
I can only comment on Darknet/YOLO. I'm not familiar enough with Ultralytics to say why it is slower, other than to guess that Python is always slower than C and C++.
This is a very constrained example, I'm interested in how this would be able to detect people in a crowd
I didn't train this network to find people and crowds. It was trained to find the things you see in the video. Like all customers who ask me to train neural networks for them, they typically want to find specific objects in machinery, or on a conveyor belt, not MSCOCO-style "find 80 random classes of things."
Good job
Definitely faster inference, but the lack of parameters hinders its accuracy harshly.
I'm guessing you didn't watch the video? Cause the accuracy is definitely higher with Darknet/YOLO.
@@StephaneCharette That is exactly my point. Does not YOLOv10n have smaller parameter size than YOLOv4-tiny?
would love to see a darknet/yolo + florence 2 + sam 2 pipeline
Great comparison.
Is there any way to run instance segmentation with yolov4 or anything open source?
What about yolo v8 tiny?
Good good stuff
is yolov4 still the best, compared to other versions?
Watch the video above and let us know what you think.
@@StephaneCharette Okay, I watched the video, but shouldn't the new versions be better to yolov4 logically? (I use Google translate, sorry for any translation errors. )
@@g-4660-i4l Darknet/YOLO with YOLOv4 is still better than YOLOv5, v6, v7, v8, v9, and now v10.
@@g-4660-i4l The version bumps are just marketing. while there are technical changes, they do not provably provide obvious speed, accuracy or training improvements.