liuyulvv

@watermellon2018 Good jobs, can you share your codes? The picture seems to be from the bgr channel.

`PhotometricInterpretation` in my dcm is `MONOCHROME1`, and I try to change it to `MONOCHROME2` but not working. Change `PresentationLUTShape` to `INVERSE` still not working. Here is my solution: ```python img...

You need to new a `holistic_landmark_only_face.pbtxt` like `holistic_landmark_cpu.pbtxt`, and remove `HandLandmark`, `PoseLandmark` can't be removed because `FaceLandmark` need its output. Why don't you try `FaceMesh`?

@xxqwf [Hand](https://google.github.io/mediapipe/solutions/hands.html)是只识别手势的，[Pose](https://google.github.io/mediapipe/solutions/pose.html)是只识别姿势的。 Holistic本来就是面部、姿势、手势的集成方案。如果你想一个包根据需求只检测某一部分的话，你可以试试将Holistic打包成一个库（我不懂安卓，在pc上是打包成链接库），然后根据需求建立自己的`graph`在运行的时候更改`graph`参数(pc命令行里面的`--calculator_graph_config_file`)，在`graph`里面去掉不需要的`calculator`和`output_stream`

@xxqwf 从[这里](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/graphs/holistic_tracking/holistic_tracking_cpu.pbtxt#L39-L48)可以看到Holistic的输出除了Hand和Pose，还有Face的。 Face的输出是在[这里](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/modules/holistic_landmark/holistic_landmark_cpu.pbtxt#L127-L146)计算的，直接注释掉，`graph`那边Face相关的[output_stream](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/graphs/holistic_tracking/holistic_tracking_cpu.pbtxt#L45)和[这里](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/graphs/holistic_tracking/holistic_tracking_cpu.pbtxt#L65)也注释掉，应该是可以的。

@xxqwf 我注释了这些地方，在PC上CPU跑起来了 [1](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/graphs/holistic_tracking/holistic_tracking_cpu.pbtxt#L45) [2](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/graphs/holistic_tracking/holistic_tracking_cpu.pbtxt#L65) [3](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/graphs/holistic_tracking/holistic_tracking_to_render_data.pbtxt#L16) [4](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/graphs/holistic_tracking/holistic_tracking_to_render_data.pbtxt#L450-L752) [5](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/graphs/holistic_tracking/holistic_tracking_to_render_data.pbtxt#L771-L772) [6](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/modules/holistic_landmark/holistic_landmark_cpu.pbtxt#L93) [7](https://github.com/liuyulvv/mediapipe/blob/e42a9ed4c9a31a10d2d1e34ccec647b6504688a1/mediapipe/modules/holistic_landmark/holistic_landmark_cpu.pbtxt#L127-L146)

@xxqwf PC上我是做到了，就是把Pose和Hand打包成动态链接库，main函数根据需求读取graph就行了

都可以吧，只要你可执行文件链接上就可以。我是打包成一个，要不然要写两个头文件接口。 > 打包成一个动态链接库，还是两个呢

@xxqwf [DLL](https://github.com/liuyulvv/mediapipe/blob/e3a6cc289078aacc73d1a4fb7c62026504cc0efc/mediapipe/dll/BUILD#L23-L24)这个是把Holistic和Iris打包在一起的，只是测试用的

In util.py: ```python def MRAE(pred:Tensor, targ:Tensor)->Rank0Tensor: "Mean Relative Absolute Error between `pred` and `targ`" pred, targ = flatten_check(pred, targ) return (torch.abs(pred - targ) / (targ + 1e-8)).mean() class MRAELoss(nn.Module): def...