Custom keypoint detection on CenterNet


I’m trying to train custom keypoint detection on 2 classes each containing 3 keypoints using CenterNet HourGlass104 Keypoints 512x512 from the model zoo. I am able to modify the pipeline and train on single class with keypoints, but I was unable to find any relevant examples that would help me out to setup the training job for 2 different classes each containing 3 keypoints.

Would appreciate any help or directions.

P.S I was using this guide to start of GitHub - prabhakar-sivanesan/Custom-keypoint-detection: Custom keypoint detection using Tensorflow object detection API

Hi @Benas_Mili,

In the first step, modify the category information. In the train_input_reader section of the configuration file, modify the label_map_path parameter to point to a label map file that contains information about your categories. In this file, add the two categories and their corresponding IDs.

in second step modify the keypoint information, In the model section of the configuration file, modify the num_keypoints parameter to 3 (since each class contains 3 keypoints). Additionally, modify the keypoint_spec parameter to include the new classes and keypoints, and train and evaluate the model after modifying these steps.Please make sure that you will need to provide the path to the modified pipeline configuration file and the checkpoint directory.

Please let me know if this helps you.