Is PP-YOLOE Better than YOLOv5 & YOLOX?
PP-YOLOE just got released but is it better than all its other YOLO counterparts, like YOLOv5 and YOLOX? Let’s find out in this article.
What is PP YOLOE?
First off what the F#%& is PP-YOLO-E? Well, PP is short for Paddle Paddle.
No, not that type of Paddle.
Paddle stands for (PArallel Distributed Deep LEarning) which is a deep learning framework developed by Baidu. You know the Chinese version of Google… Yeah, that one.
So Baidu, started out creating their own flavour of the YOLO Objector Detector that is based on the already popular YOLOv3 for which they branded it as PP-YOLO. If by now, you are scratching your head wondering WTF is YOLO, then I assume you know nothing John Snow and that you should watch this video explaining what YOLO is, in detail.
Now before we continue, I just want to say that I’m so glad that they called it PP-YOLO rather than the next version of the main YOLO like YOLOv5, which was surrounded by a whole lot of controversy (controversy documentary right up here). And lastly, if you want the full buffet of all-you-can-eat YOLO courses, then check out Augmented Startups for all types of YOLO that you think you can stomach, like YOLOR, YOLOX, YOLOv4, and more.
Anyways moving on… the first iteration of PP-YOLO was better than YOLOv4 in both Mean Average Precision(mAP) and inference time on the COCO dataset and tested with the Nvidia V100 GPU.
In April 2021, Baidu released their second iteration of PP-YOLO, called PP-YOLOv2… surprise surprise. According to their paper called PP-YOLOv2 — A practical Object Detector, they have mentioned that they had surpassed existing object detectors like YOLOv4-CSP and YOLOv5-l with the same amount of parameters. Okay, okay, competition is getting heated now.
In some cases, they mention that they had achieved comparable performance while still being 15.9% faster than the Yolov5X model.
This brings us to PP-YOLOE. Now, I don’t know whether it was the pandemic or a change of management. But why in the name of Chuck Noris, did they call it YOLOE? Goddammit! Could they not have called it something like PP-YOLOv3.
Was it because of all the newer models being called YOLOX, YOLOR, YOLOP, YOLO-FU? AAAhh!?
The only reasonable explanation is that maybe PP-YOLOv3 would have been confused with YOLOv3, in terms of search engine optimization. I don’t know honestly.
Anyways on the 2nd of April 2022, Baidu released their subsequent paper PP-YOLOE — An Evolved version of YOLO, which is said to be an industrial state-of-the-art object detector with high performance and friendly deployment.
They state that they were inspired by YOLOX for its anchor-free method equipped with dynamic label assignment to improve the performance of the detector, which they claim has significantly outperformed YOLOv5 in terms of precision. Knowing this, they further optimised their previous work in PP-YOLOV2.
Specifically, PP-YOLOE achieves 51.4 mAP on COCO with 640 × 640 resolution at the speed of 78.1 FPS, surpassing PP-YOLOv2 by 1.9% AP and YOLOX-l by 1.3% AP. Moreover, PP-YOLOE has a series of models (small, medium, large), similar to Mac Donalds.. cough I mean YOLOv5.
Okay ladies and gents, so is PP-YOLOE the winner right? Well not quite, while it is shown to be a high-performance detector, there are still no comparisons with other models like YOLOR. That’s a match-up that I would like to see. In comparison to PP-YOLOE, I anticipate that YOLOR would be more accurate but slower in terms of performance (inference time).
The other thing to consider is the ease of use and community support. YOLOv5 has around 26k Github stars, compared to 6.3k stars for YOLOX and around 7.6k Github stars for PP-YOLO, which has some serious catching up to do. Models like YOLOv5 are versatile in a way that you can deploy them to smartphones and developers have reported that it’s quite easy to use. So it all depends on your application which would dictate the selection of the model.
Now, let me tell you a secret, which is YOLO, You Only LIKE Once, meaning that you can only LIKE this article ONCE… not twice (otherwise it will unlike this article). So please like this article… Once Please. And if you want to learn more about PP-YOLOE architecture and implementation, then tell me in the comments down below.
Subscribe to the Augmented Startups YouTube Channel and based on the demand we’ll create a tutorial series on this.