Ultralytics Releases YOLOv8
Created on January 14|Last edited on January 14
Comment
Ultralytics just released YOLOv8! It can be found here. It's hasn't even been a year since I last used Ultralytics' YOLOv5 for an object detection task and here we are at v8! Machine learning open-source development and research move insanely fast and every new day is an exciting mystery. As for the actual code and repository, click here. They also have a webpage!
A little background on YOLO (You Only Look Once). Prior to the first YOLO release (or v1) by Joseph Redmon et al., R-CNNs existed. In 2013, Ross Girshick et al. proposed Regional-Convolutional Neural Networks or R-CNNs for short. This architecture featured a CNN paired with a selective search algorithm to generate bounding boxes around regions within an image that an object might likely reside in. These proposals, as they are called, are fed into a CNN. This multi-stage pipeline, termed an R-CNN, was the go-to for object detection. It later evolved into another paper Fast R-CNN, then eventually Faster R-CNN. The changes made in these subsequent papers featured replacing this selective search algorithm with a neural network such that the entire pipeline or R-CNN could be trained end-to-end. The multi-component aspect of the R-CNN was a lot to manage and that's where YOLO comes in.
As the name suggests, YOLO only requires 1 forward pass of an image or, simply put, the model only needs to see the image once. In R-CNNs, the image is broken down into multiple proposals and these are all fed into another neural network. You can imagine, that for a single image, that neural network would have to parse through hundreds of said proposals. YOLO looked once and was trainable end-to-end. Fast forward to today, YOLO underwent multiple improvements from v1 up to v5 with v1 through v3 led by Joseph Redmon, and the later versions carried on by other developers. Ultralytics introduced YOLOv5 which was available for the community to play with all via command line statements! There were also an assortment of other YOLO flavors like YOLOR and YOLOX. Now with v8, the interface is much more customizable (you can interact with YOLOv8 via command line or through Python code). Additionally, the model performs, as you guessed, even better!
Besides the immediate benefits of this release —better performance and features, I could imagine a tool like this being a staple for any computer vision engineer or scientist. For one, I know Kagglers would benefit greatly in their computer vision competitions with this new toolkit! :)
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.