YOLOv7 Pose vs. Mediapipe! Human Pose Estimation Benchmark and which one to choose? Hey there, welcome to LearnOpenCV. YOLOv7 pose is the new kid on the block for Human Pose Estimation. We will pit it against the popular Mediapipe pose model using challenging videos and share our findings through this video. I am sure this will help you choose the right pose estimation model for your next project. Before we start, let's do a quick refresher on Human Pose Estimation for those who are new to this topic. If you want to jump into the comparison section, skip to the fourth minute of the video. Now, what is Human Pose Estimation? Human Pose Estimation is the task of predicting the location of major joints of the human body. It is also referred to as Keypoint Detection. Now you may ask how many points? There are many datasets for Keypoint Detection with different topologies.

The most common being the Coco dataset, which has 17 keypoints. Others include the MPII dataset, which has 16 keypoints, the AI Challenger dataset, and the CrowdPose dataset, both with 14 keypoints, and so on. Human Pose Estimation has possible applications in many industries. For example, you can build AI fitness trainers like that of peloton.ai. It can be used in healthcare services, sports analytics, activity recognition, surveillance, in the VR industry for gaming, and human-computer interaction. There has been extensive research on this topic, and many algorithms have been proposed over the years. Some of the popular ones are OpenPose, AlphaPose, HRNet, DEKR, BlazePose, and YOLO Pose. Now, what is YOLO Pose? YOLO Pose is an end-to-end pose estimation model. To understand what it means, let us quickly discuss the two general approaches for pose estimation. In the Top-down Approach, we first detect persons in the image and then use these

boxes to find key points. As you can guess, this would give pretty good results but would be too slow if there are too many people in the scene. The Bottom-up Approach uses heat maps to find key points and then uses a grouping algorithm to map the key points to each person. This is a single-stage and thus it is fast, but you can imagine it won't be very accurate in case of crowded scenes. YOLO Pose moves away from both these approaches and directly optimizes the object keypoint similarity metric, which jointly detects the key points as well as the bounding boxes. So, it does not need the grouping step done in Bottom-up Approaches, and because the key points and boxes are predicted in a single inference step, it does not compromise on speed as well. Thus, it brings in the best of both approaches. YOLO Pose is designed with a general architecture so that you can plug in any object detection model. The original paper was based on YOLOv5.

Now since the architecture of YOLO Pose is general, the YOLOv7 authors adapted it to YOLOv7 and hence the name YOLOv7 Pose. We want to test YOLOv7 Pose in terms of speed and accuracy and compare the features with Mediapipe as it is a very accurate and real-time model. Mediapipe is a framework that provides multiple computer vision solutions and one of them is pose estimation. It uses the popular BlazePose model for pose estimation. Now let's see the differences between YOLOv7 Pose and Mediapipe. YOLOv7 is a multi-person pose estimation model whereas Mediapipe is for single-person pose estimation only. YOLOv7 has 17 keypoints as compared to 33 keypoints for Mediapipe. The default input size of YOLOv7 is 960P whereas for Mediapipe it is 256 by 256. This means it is trained for small image sizes as well.

YOLOv7 performs detection on every frame, whereas the Mediapipe framework uses a detection plus tracking approach so as to increase the speed as well as provide a more stable output, reducing the jitter. Mediapipe is optimized for CPU and Edge devices, but there is no such optimization done for YOLOv7 Pose. Now, let us see how they compare against each other by running them on some challenging videos. First, we see the results that we get out of the box with the default values. Here, we can see the drastic difference in FPS between Mediapipe and YOLO. This is because the input size of Mediapipe is 256 by 256, whereas for YOLO, it is 960P. This comparison is a little unfair. So, in our next run, we change the input size of YOLO to 256P and get these results. It gives an 8x boost to YOLO but still lags behind Mediapipe in terms of FPS. The next thing

we want to check the amount of jitter or flickering present in the models. For this to be prominent, we zoom in on a particular area of the video and see the results. The difference is evident that there is a lot of flickering in the YOLO results as compared to Mediapipe. Even for the default 960P input size, the YOLO Pose model flickers more than the Mediapipe model. The next comparison is checking for scale variations, like what happens when the person's size varies. We can see that the YOLO model with a small input size of 256 fails to detect the person, but Mediapipe is able to detect the key points. YOLOv7 with a larger input size of 960 detects for some frames but then fails, but Mediapipe continues to detect as it performs tracking. So, if it has detected the person once, it is able to track the key points across frames.

The other common challenge in pose estimation is occlusion. In this specific example of occlusion, Mediapipe is mistaking the horse's legs for human, but YOLOv7 is able to detect the occluded legs as well. But this seemed very uncharacteristic of Mediapipe, so we tested it on some more videos. In this video, we can see that both the models are performing decently well even when only a part of the person is visible. Still, on closer observation, YOLO outputs look better than Mediapipe for occlusion. In case of low lighting conditions, Mediapipe makes more mistakes as compared to YOLOv7, but it is still a close call. Finally, it's time to put them to the extreme test. This video is difficult because the motion is fast and the person is upside down. The YOLOv7 model is not able to keep track of the fact that the person is upside down and is still predicting the points as

if the person was upright. Notice that the blue lines are for the arms and orange lines are for the legs. Mediapipe completely outperforms YOLO in this video for both the input sizes. Now let's summarize the results and look at the key takeaways. Mediapipe cannot be used for multi-person pose estimation, and the discussion ends there if you are looking for a multi-person pose estimation model because Mediapipe is only for single-person pose estimation. Mediapipe has less jitter as compared to YOLO. This can be attributed to the additional tracking step. If you have a GPU and you want very accurate results, use YOLOv7 with a larger image size. If you want a model capable of running at real-time on the CPU, go for Mediapipe. When you need to perform pose estimation on still images, Mediapipe might not be the best option since a big advantage of Mediapipe comes from the detection plus tracking framework. Note

that it has a separate functionality for running pose estimation on images as well. What do you think about the results? We would love to hear your observations as well. I hope you found the video interesting. Do LIKE and SHARE the video and SUBSCRIBE to our Channel. Until next time!

Official YOLOv7 Pose vs MediaPipe | Full comparison of real-time Pose Estimation | Which is Faster?

Official YOLOv7 Pose vs MediaPipe | Full comparison of real-time Pose Estimation | Which is Faster?