How to obtain the center point coordinates and depth values during NYU inference about a2j HOT 4 OPEN

Hello, I don't understand the purpose of the central coordinates. What else is there besides getting bbox

from a2j.

@weilanShi
Hi, the center point of public dataset is taken from V2V code (https://github.com/mks0601/V2V-PoseNet_RELEASE), and if you have to test on your own data, you may need to train a center point regressor as in V2V, or you can train a object detector that predicts bounding box first.

from a2j.

@weilanShi Hi, the center point of public dataset is taken from V2V code (https://github.com/mks0601/V2V-PoseNet_RELEASE), and if you have to test on your own data, you may need to train a center point regressor as in V2V, or you can train a object detector that predicts bounding box first.

非常感谢您的回复, 我最终是想去除center-point 这个需要提前准备好的值(主要是不想推理时要太多模型), 我看代码里center-point 的作用主要有两个: 1. 用来结合xy_thres 取出大致的bbox区域；2. 用来结合depth_thres去除其异地点,同时将 imgResize-center(为了将imgResize的x,y,z统一到同一个量级). 为了去除center-point 这个需要提前准备好的值,所以我想用以下方法替代center 的两个作用: 1. 使用手部检测器来获取bbox；2. 通过取imgResize的中心小roi区域的depth_mean 来作为 center 值,同时重新使用该方法得到mean/std. 按照上述方法进行了一些实验,目前loss比艰难收敛. test error为24. 想请问下上述方法是否有问题, 或者您能给出其他一些建议.非常感谢,期待回复,提前祝国庆节快乐!

from a2j.

@weilanShi Hi, the center point of public dataset is taken from V2V code (https://github.com/mks0601/V2V-PoseNet_RELEASE), and if you have to test on your own data, you may need to train a center point regressor as in V2V, or you can train a object detector that predicts bounding box first.

你好,我还有一个问题: UVD坐标里的z就是采集数据时,当时手部关节点到深度相机的距离吧? 还是需要进行其他坐标转换?

from a2j.