ashar6194/3d_human_pose_estimation

Semantic Segmentation Pipeline for Range Images for Identification of 3D Pose

SegNet- UBC Implementation

This is modified implementation of Segnet inspired by segmentation network proposed by Kendall et al. on UBC Human Pose dataset and Berkely MHAD.

Implementations of U-Net, Pointnets(++) are next in line for the whole task of pose estimation.

The I/O pipelines for both the datasets are ready.

In the config.py file, the dataset_name needs to match the data directories you create in your input folder. You can use UBC_easy and segnet-32.

Train and test

Generate your TFRecords using tfrecorder.py. Make sure the dataset is downloaded in and make it compatible with the directories mentioned in the tfrecorder.py file.

Once you have your TFRecords, train SegNet with python train.py. Analogously, test it with python test.py.

Also, the regression pipeline for both of the above datasets are written in MATLAB, and will be up in a separate repo shortly. In case of any questions, drop me an email at [email protected]