Results 15 comments of Tomomas

@DaChaoXc Tensorrt使用的是对称量化,就算是非对称量化,也是有zeropoint,不存在你说的仅对正半轴取值

![image](https://user-images.githubusercontent.com/30460509/95952897-8dc7db80-0e2b-11eb-8210-7cff00e15b5c.png)

when will add this method?

@guker The question be solved. 2D heatmap represent x,y image coordinate. 3D heatmap use x,y,x directly. and than 3D position use 2D message.

@sameeroor @moranli19 Mano_Scale means proportion between Z_{id=9,camera space} and focal length from Mano files. However another scale indicate length of idex=[9,10] from scale files. About Mano scale, you need see...

@lzneu The idea of this project that boundary point can easily to predict boundary. Therefore. set 0.1 can predict accuracy more than 0.5. thank you for author to share this...

@cuimiao187561 Firstly try to increase trunc_threshold paramter. Secondly try finetune net by author provide method. finally you can modify nms strategy such as supplement by manual.

how to run with audio driven?

@dollarser 直接将onnx-fp32转化为onnx-int8即可(ORT推理,官方有demo),和onnx-fp32转trt-int8类似的。

Also get this error in py3.9.12+MNN2.0.0