In recent weeks we have got some piece of surprising news in the field of computer vision. The YOLO(You Only Look Once) series got a new member named MT-YOLOv6 which can also be called YOLOv6.
YOLO series models are well known for real-time object detection and these all models are being developed by the Ultralystics. Update by update we can see that they are enhancing the speed and accuracy of the procedure. The development of YOLOv6 took place at the Vision Intelligence Department of Meituan and one of the interesting things about the model is that it is available to everyone as an open-source. The technology team of Meituan introduced their model as YOLOv6 because they took inspiration from the original YOLO series. Let’s take a look at the comparison between the new and the older versions of YOLO.
How does MT-YOLOv6 compare to YOLOv5?
According to the research team the YOLOv6 has outperformed other YOLO models like YOLOv5 in terms of prediction accuracy and prediction speed. They have tested this model using the COCO dataset. This model is supporting various deployment platforms helping is simplifying the deployment work. Below the images is the representation of this comparison(taken from GitHub).
Here above on the right side, we can see the graph between the accuracy percentage of different models including YOLOv6 and frames processed per second while using the COCO dataset. On the left, we can see the accuracy given by the models while they are processing only one image.
In terms of development, we can consider this model as the straight-up-gradation of YOLOv5. Below are some improvements that the team has performed:
- Uniformly designed backbone and neck of the system so that they can be more efficient.
- Enhanced the effectiveness of decoupled head of the network using optimisation techniques.
- They used an anchor-free paradigm of training while the program is getting supplemented by the SimOTA strategy of data labelling and the SloU strategy of applying bounding boxes to improve the accuracy of detection.
As discussed before, this model is open-sourced and using the link we can access the codes where pre-trained weights for nano, tiny, and small model sizes are also available.
References
- Article “YOLOv6: A fast and accurate target detection framework is open source” published by Meituan technical team on June 23, 2022
About DSW
Data Science Wizards (DSW) is an Artificial Intelligence and Data Science start-up that primarily offers platforms, solutions, and services for making use of data as a strategy through AI and data analytics solutions and consulting services to help enterprises in data-driven decisions.
DSW’s flagship platform UnifyAI is an end-to-end AI-enabled platform for enterprise customers to build, deploy, manage, and publish their AI models. UnifyAI helps you to build your business use case by leveraging AI capabilities and improving analytics outcomes.
Connect us at contact@datasciencewizards.ai and visit us at www.datasciencewizards.ai
Thank you for reading!