MT-YOLOv6: A YOLO-Inspired Object Detection Model Released
MT-YOLOv6 has entered the YOLOv5 spinoff scene with surprising speed and accuracy benchmarks.
Created on June 27|Last edited on June 27
Comment
YOLOv5 (You Only Look Once) is a fairly well known model made by Ultralytics for one-stage real-time object detection in videos and images, applicable for mass object detection use cases such as parsing dense traffic or grocery store aisle product detection. The model has also spawned a number of copycats and offshoot models, one of which we're highlighting today.
MT-YOLOv6 (or just YOLOv6 for short) is a newly released model which was inspired by Ultralytics' YOLO models and is developed by researchers at Meituan. All the code needed to get this model running is available in the GitHub repository, including pre-trained weights for nano, tiny, and small model sizes (4.3M, 15M, and 17.2M parameters respectively) - larger model sizes are coming soon.
How does MT-YOLOv6 compare to YOLOv5?
According to benchmarking performed by the team, YOLOv6 outperforms YOLOv5 and other YOLO models in both accuracy and speed on the COCO dataset. The following image (taken from the GitHub) shows:
- Right: COCO accuracy percentage [y] vs frames processed per second [x] with each point being various sizes of each model (small/tiny/nano).
- Left: COCO accuracy percentage [y] vs milliseconds taken to process a single image [x] for various image resolutions (384/448/512/576/640).
In some Twitter and YouTube discussions, observers find that YOLOv6 isn't simply a straight upgrade compared to YOLOv5. Some things to pay attention to in the following video and in other comparison videos is that while YOLOv6 is often able to detect smaller object more reliably, it overall flickers quite a lot compared to YOLOv5 and also seems to struggle more with close-up objects.
There's a notable stability that YOLOv6 lacks compared to YOLOv5, but it's improvements with small object detection in densely packed scenarios is impressive. Pay attention to how many more individual cars are detected, especially along the bridge and raised streets at the side, in this comparison video:
Ultimately, each model has it's strengths and weaknesses, but YOLOv6 is definitely a nice addition to the existing lineup of YOLO spinoffs.
Find out more
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.