We have introduced a lot of research on using RGB cameras for motion capture. Most of them simulate bones for tracking. A few of them use a rough model with the simulated bones. Recently, Facebook AI Reaserch (FAIR) open sourced a technology that maps all human pixels of 2D RGB images to 3D models in real time-DensePose, and it does not use the bone tracking we often introduce, but a very intensive Project tracking to build 3D models. They also perform well outdoors and in loose clothing. Also supports simultaneous tracking of multiple people. How to understand this intensive? For general bone tracking, most of the tracking points are between ten and twenty, and no more practical effect. The DensePose tracked a total of 336 points, densely covered with dots. (Intensive phobia withdrawal) The reason for tracking so many points is the data necessary to build a smooth and smooth 3D model. The hard work is also worth it. DensePose performs well both outdoors and in multiple people, and can change the clothes of people in the scene in real time. Then take a look at how they did it. In order to allow the machine to learn, the researchers manually marked 336 points in 50,000 photos. This step alone is a huge project. If you mark the annotations step by step, I do n’t know when it will be completed. The researchers split a person into 24 parts, namely the head, upper torso, lower torso, upper arm, lower arm, thigh, lower leg, hand, and foot. Each section marks 14 points. The head, hands, and feet are manually marked by the person. At the same time, the annotator is also required to mark the areas covered by clothing, such as loose skirts, when marking. After these tasks are completed, they enter the second stage. The researchers sample each unfolded area and obtain 6 different marker maps, providing a two-dimensional coordinate map to allow the marker to more intuitively determine which marker is correct. Finally, the planes are reassembled into a 3D model, and the final calibration is performed. With these two steps, the researchers were able to obtain accurate marks efficiently and accurately. However, there are large errors in the torso, back and hips. The next step is the stage of deep learning. At this time, a good solution is like a catalyst with superior performance. The researchers adopted a method similar to DenseReg of the Mask-RCNN architecture to form the 'DensePose-RCNN' system, and further developed it to improve the accuracy of training. First, the appearance of the pixel is roughly estimated, and then it is aligned with the exact coordinates. The key point branch of DenseReg MaskRCNN uses the same architecture, consisting of 8 alternating 3 × 3 full convolutions and 512 channels of ReLU layers. Thanks to Caffe2, the resulting architecture is actually as fast as Mask-RCNN. In order to reduce the error rate, a "teacher" network was also trained to reconstruct the ground, deploy its complete image domain, and generate a dense supervision signal. The researchers compared the semi-automatic supervision of human supervisors with the "teacher network", and the result was that the "teacher" won. The researchers also compared their methods with SMPLify. In terms of model simulation, the researchers' bottom-up feed-forward method greatly outperformed the iterative model fitting results. At the same time, FCN is significantly worse than 'DensePose-RCNN when dealing with multiple people, and its advantages are also very prominent when compared with other schemes. Finally, as shown at the beginning, the overall effect can handle a large number of occlusions, successfully simulating the people behind the clothes, but there is one thing to note that everyone is fitted by a fixed curvature. And it performs very well in multiplayer situations. customized high voltage,High voltage ESS, HV battery, large scale battery; one-stop solution provider for energy storage; high voltage battery, high voltage battery storage system, high voltage battery cabinet Shenzhen Enershare Technology Co.,Ltd , https://www.enersharepower.com
