Training PoseClassificationNet for custom dataset

• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc) PoseClassificationNet

PoseClassificationNet training needs data.npy & label.pkl files for train, test and validation.

I have created json files using deepstream-bodypose-3d app for each short video clips (<300 frames).
Then pkl file is created as samples names and class ID as discussed here in similar format as follow.

[["xl6vmD0XBS0.json", "OkLnSMGCWSw.json", "IBopZFDKfYk.json", "HpoFylcrYT4.json", "mlAtn_zi0bY.json", ...], [235, 388, 326, 306, 105, ...]]

The issue is npy file to create for train, test and valid.
Downloaded Nvidia datasset and checked data inside npy files.
For train_data.npy, shape of numpy array is (9441, 3, 300, 34, 1).

(1)How to consolidate data inside json files to have such numpy array. Any script for producing numpy file.
(2)Each video clip need to have 300 images exactly?
(3)Data inside Json file is not normalized.
But data inside numpy array are normalized. How to do normalization?
(4) Which graph_strategy to use in training?. Json files are created using deepstream-bodypose-3d

Then in (9441, 3, 300, 34, 1), 3 represents number of channels.
Json data is meant for locations of joints in image.
How data in Json be separated for 3 channels?

I have to follow this command from Converting The Pose Data session.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.