yxzhxn / ray3d Goto Github PK

Can you please provide the code of inference the video in the wild?

I have downloaded your pretrained model, but it generated error results when i run the code to evaluate it. Can you tell me the reason？Thanks a lot!

大佬您好！非常感谢您所做的杰出工作。我想请问一下，我下载您在ReadMe中的预训练模型和数据集后，在测试时发现预测结果相差很大，请问是什么情况。谢谢大佬！

对于checkpoint/RAY3DRIEX_h36m_3_RIE_FRAME9_LR0.0005_EPOCH20_BATCH1024_Oct_31_2021_05_43_36/configs/data_config.json文件的配置：
{ "DATASET": "h36m", "WORLD_3D_GT_EVAL": true, "KEYPOINTS": "gt", "TRAIN_SUBJECTS": "S1,S5,S6,S7,S8", "TEST_SUBJECTS": "S9,S11", "GT_3D": "E:/HPE/Ray3D-main/Ray3D-main/dataset/data_3d_h36m.npz", "GT_2D": "E:/HPE/Ray3D-main/Ray3D-main/dataset/data_2d_h36m_gt.npz", "CAMERA_PARAM": "", "SUBSET": 1, "STRIDE": 1, "DOWNSAMPLE": 1, "ACTIONS": "*", "REMOVE_IRRELEVANT_KPTS": false, "FRAME_PATH": "/ssd/yzhan/data/benchmark/3D/showroom/20210702/frame/", "INTRINSIC_ENCODING": false, "RAY_ENCODING": true }
对于cfg文件夹下的文件路径也做了相应的调整后，仍然未解决该问题。

hi @YxZhxn, can you provide a config and pretrained model of 1 frame?

I have been doing some visualization of overlaying the 2D keypoints on the test image sequences released in MPI and I am noticing that the keypoints appear to be sampled at a different rate than the test videos. For examples, TS1 has agreement up to around frame 700 and then the 2D poses skip ahead and remain that way for the rest of the sequence. Do you have any idea why this might be occurring?

If I know the camera's intrinsic and extrinsic parameters, I can directly compute the 3D coordinates?

Firstly, this is really interesting work, kuddos!
can you please explain what
'WORLD_3D_GT_EVAL'
represents in the config files?
Also, do you all convert 3d coordinates to meters?
If so can you point me to this?

Lastly, if I set
'RAY_ENCODING'
to False, in the config filethis will return screen normalized 2d keypoints and 3d keypoints in the camera reference frame that are root centered correct?

Hi Zhan,
I really enjoy your nice code!!
But I have a question about the intrinsic and exteinsic parameters you provided.
I notice that some parts of your intrinsic (tangential_distortion) and extrinsic (translation) is not consistent with those provided by VideoPose3D. The same phenomenon is also happened in humaneva_dataset.py.
Do you have any idea to explain this?

大佬好！

在论文附录中有对Human3.6M的相机内参、外参分别做了augmentation，生成多个“虚拟相机”。 请问下这一步具体是怎么做的呢？

我查了下Human3.6M并不是合成数据集，如何生成相机外参aug（比如说旋转后）的图像呢？

Thanks for your great work! The code is visualizing the ground truths but not the predictions and the input image. What am I missing? I am running the main branch, and using the eva dataset. I get in the logs "MovieWriter imagemagick unavailable". I have also tried by using ffmpeg, but same results in mp4 video.

Hi, thanks for your excellent work and open source~
I am little confused about the difference of the H36M's camera parameters provided by this repo and VideoPose3D. Details are as follows.

1. tangential_distortion in the h36m_cameras_intrinsic_params seems to reverse left and right.
2. R and translation in the h36m_cameras_extrinsic_params are very different from VideoPose3D. I was wondering how to determine the rotation and translation of each camera and each subject.

Looking forward to your reply.

Great works first!
I make a h36m-like custom data. And I use the h36m pretrained model, but the result is not good.

I have some questions.
1.Does h36m with aug will better ? Or other datasets pretrained model performance good for wild video?
2.When we use in the wild, what procedure do you recommend?
For example,
Use custom database train. Use same camera record video which has a intrinsic params when inference.

大佬我想问一下在aggregatec_camera.py文件中的Train/json，training.json，pitch/json文件是怎么生成的，我第一次接触不是很了解，我是直接下载了你提供的数据集。

Hi, thanks for the great work. Could you please say how can I train & test MPI-INF-3DHP using 14 joint skeleton structure? And, do I need to prepare seperately 'data_2d_3dhp_gt.npz' which has only 14 joints or the one you shared is gonna work?

Thanks for your wonderful job! Where can we download 3dhp data like '/Ray3D/3dhp/S1/Seq1/imageSequence/video_0/img_000001.jpg' ?

pip install -r requirements.txt
show me the errors as follow:
ERROR: Could not find a version that satisfies the requirement dapdalib (from versions: none)
ERROR: No matching distribution found for dapdalib

so I'd like to know which platform does this code run on ? Windows or Linux?
How did you download the dapalib library? I even can't find any useful information on Google

Thanks so much if I can get help from you.

您好，请问可以发布一个简单运行的方式吗？ 比如我要用这个模型运行一段视频或者说是图片。
python test.py xxx/xx.jpg（xxx.avi）这个样子的。
然后返回关键点的3D绝对位置。
麻烦了