Predictions Problems

[aIacoella]

Hi, I am trying to run Shapy locally on my computer.

After finishing the installation, I managed to run shapy regressor on the repo's sample data.

Unfortunately there seems to be some issue with the predictions as they all look the same and don't resemble the input data.

The input data is the one provided in the repo (images and openpose).

The output of the regressor all look weird:

img_00.jpg

img_00_hd_stage_02_cat.png

img_01.jpg

img_01_hd_stage_02_cat.png

img_02.jpg

img_02_hd_stage_02_cat.png

The issue is even more clear when looking at the output from the virtual measurements:

Processing: img_00.npz
Virtual measurements: mass: 63.45 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_01.npz
Virtual measurements: mass: 63.55 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_02.npz
Virtual measurements: mass: 63.33 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_03.npz
Virtual measurements: mass: 63.27 kg height: 1.69 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_04.npz
Virtual measurements: mass: 63.24 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_05.npz
Virtual measurements: mass: 63.18 kg height: 1.69 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_06.npz
Virtual measurements: mass: 62.90 kg height: 1.69 m chest: 0.94 m waist: 0.80 m hips: 0.93 m
Processing: img_07.npz
Virtual measurements: mass: 62.98 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.93 m
Processing: img_08.npz
Virtual measurements: mass: 62.64 kg height: 1.69 m chest: 0.94 m waist: 0.80 m hips: 0.93 m
Processing: img_09.npz
Virtual measurements: mass: 63.78 kg height: 1.70 m chest: 0.94 m waist: 0.81 m hips: 0.94 m
Processing: img_10.npz
Virtual measurements: mass: 63.20 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_11.npz
Virtual measurements: mass: 63.13 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_12.npz
Virtual measurements: mass: 63.32 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_13.npz
Virtual measurements: mass: 63.04 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_14.npz
Virtual measurements: mass: 63.11 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_15.npz
Virtual measurements: mass: 62.85 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.93 m
Processing: img_16.npz
Virtual measurements: mass: 62.93 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.93 m
Processing: img_17.npz
Virtual measurements: mass: 63.04 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_18.npz
Virtual measurements: mass: 62.89 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.93 m
Processing: img_19.npz
Virtual measurements: mass: 62.99 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_20.npz
Virtual measurements: mass: 63.10 kg height: 1.69 m chest: 0.94 m waist: 0.80 m hips: 0.94 m
Processing: img_21.npz
Virtual measurements: mass: 63.02 kg height: 1.70 m chest: 0.94 m waist: 0.80 m hips: 0.93 m

---

Does anyone know why something like this might happen? As I said I've only followed the installation steps and run the model.

[yunkchen]

you could check your pose json files in openpose.

[Liyuanhong233]

你好，请问下openpose 中的姿势 json 文件是怎么生成的呢？

[yunkchen]

你好，请问下openpose 中的姿势 json 文件是怎么生成的呢？

类似这样，用openpose推理后，结果format成shapy的输入格式

people_data_tmp = {
                "person_id": [-1],
                "pose_keypoints_2d": datum.poseKeypoints[p_idx].flatten().tolist(),
                "face_keypoints_2d": datum.faceKeypoints[p_idx].flatten().tolist(),
                "hand_left_keypoints_2d": datum.handKeypoints[0][p_idx].flatten().tolist(),
                "hand_right_keypoints_2d": datum.handKeypoints[1][p_idx].flatten().tolist(),
                "pose_keypoints_3d": [],
                "face_keypoints_3d": [],
                "hand_left_keypoints_3d": [],
                "hand_right_keypoints_3d": []
            }

[Liyuanhong233]

你好，请问你用的是官网的openpose识别的吗，我编译安装完成之后识别的人体都没有关键点，查了很多原因都没有解决，不知道是怎么回事，另外我用了一个非官网的代码pytorch-openpose，识别出来的结果只有18个人体关键点，推理之后生成的3D人体不准确，是不是在生成3D人体的时候还需要脸部，手部的关键点呢？

[yunkchen]

你好，请问你用的是官网的openpose识别的吗，我编译安装完成之后识别的人体都没有关键点，查了很多原因都没有解决，不知道是怎么回事，另外我用了一个非官网的代码pytorch-openpose，识别出来的结果只有18个人体关键点，推理之后生成的3D人体不准确，是不是在生成3D人体的时候还需要脸部，手部的关键点呢？

是官网的openpose，有几种模型要选择137（好像？）关键点的，上面我代码里的datum就是openpose的返回

[Liyuanhong233]

那三种模型我都试了，都没有出结果，您可以把你的openpose工程发我一份吗，我想看看是我代码的问题，还是没编译好的问题，[email protected]，这是我的邮箱，如果方便的话，还请麻烦你发我一份，感谢！！

[yunkchen]

那三种模型我都试了，都没有出结果，您可以把你的openpose工程发我一份吗，我想看看是我代码的问题，还是没编译好的问题，[email protected]，这是我的邮箱，如果方便的话，还请麻烦你发我一份，感谢！！

https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/03_python_api.md#testing-and-developing