Hey Stephane, first of all thank you for all your materials. These are extremely helpful and to the point. Is OBB (oriented bounding box) feasible with darknet yolo? Are there any forks supporting that by chance? I am aware of yolov5-obb and fairly recently yolov8-obb but these are clearly not darknet as your yolo is.
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."
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.
@@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. )
Perhaps I need to take another look at Darknet... For a particular dataset I have, it didn't do that good and I doing YOLOv6 to be the best. Even compared to v10. SOTA doesn't mean SOTA for your dataset. It is quick and easy to experiment with the various YOLOs... BTW, the default image size is 640 x 640.
@@ckpioo Ubuntu 20.04, 24 GiB vram, 32 GB ram, AMD Ryzen 9 @ 2200 MHz. Doesn't matter, the point is not to actually train in 85 seconds, the point was to show that if your network is limited to a specific situation, you can train in a very short amount of time instead of "days" and with thousands of images.
@@leo1722467 fiz mano tá nos meus vídeos reconhecimento de armazenamento de fogo, mais depois eu parei, eu tava ou tô ainda até hoje no grupo dele no Discord,aí vi eles comentar sobre uma técnica para girar o retângulo de reconhecimento era augo novo que ainda tava lançando, aí na época eu parei , esperando essas coisas aparecer e não voltei até hoje.
This is really helpful for piglets that is still in their mothers. It can potentially prevent the piglets getting crushed by the mother when she lays down.
See the video description for details. This is part of the DarkHelp library for Darknet/YOLO. DarkHelp is at the usual location: github.com/stephanecharette/DarkHelp
Thank you for the videos. I have seen one post from reddit about darknet/yolov4 that this networks are greats and now I have really good results. Thanks!!!!
Nice, once i used edge detections to make the calculation of the rotation. But, I'm passing a image thru the NN, it's YOLO, so it's already looked for it and i think it will costs less for cpu to detect, calculate and rotate. Nice approach.
Thank you for all the dark resources, it helps me a lot. And is there have any clue on how to transplant a model to an embedded device such as ARM SBC board (ARM9 or Cortex A7), or even a Microcontroller (ESP32), etc.
I have used Darknet/YOLO on ARM processor devices such as Beaglebone, RPI, and NVIDIA Jetson devices. I doubt you'd get it to run on devices such as ESP32 with only 320 KiB of ram. Darknet requires a full operating system, and the weights alone are 25 MiB or more in size.
If a 16x16 object size (1/26th) for 416x416 network configuration is ideal (minimal) size, can I then assume that the smallest object size must be not smaller than 1/26th the size of the network?
@@StephaneCharette Appreciate it, however, I just want to report the above assumption to a client (if true). They would like to know what is minimal size of object given a particular network size as a starting point. I may later delve into using other techniques to improve small bject detection performance.
@@sokhibtukhtaev9693 I've already answered you. The minimum I've seen from Darknet with DarkHelp and tiling is 7x7 pixels. Doesn't matter if the frame is 416x416 or 9999x9999. As I've stated in other places, the minimum I personally would recommend in high-contrast situations is approx. 12x12.
@@StephaneCharette Thank you will give it a go, so far all tutorials I can find are Linux or Windows, not seen anything for Mac, are you aware of any tutorials?
See my other videos where this is discussed. For example, ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-BcC5kDNX510.html and ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-d8baNNR2EyQ.html
