What is the best way to train TensorFlow for custom keypoint tracking that can work on the web?
Right now I’m using CenterNet MobileNetV2 FPN Keypoints 512x512 to train, but the outcome is not good enough keypoints confidence is significantly less approx 30%, but the bounding box is fine. So is there any way I can improve the model confidence for keypoints?
Config :
steps 25000
epoch 12
learning rate 0.01
train dataset 1280
test dataset 319
There were no tangible guide to train a key point detection model on custom dataset other than human pose or facial key points.
Tensorflow object detection api itself provides an example python script to generate TFRecord for coco based annotations. But the script is primarily written for coco dataset which contains human pose keypoints. So with few changes to it, we can use it for any custom dataset.
To do that, edit the _COCO_KEYPOINTS_NAMES list in line no 87 with our keypoints data with the same order it appears on the annotation file. This is an important step which needs to be verified.
[quote=“Rodriguez2121, post:5, topic:13155, full:true”]
Tensorflow object detection api itself provides an example python script to generate TFRecord for coco based annotations. But the script is primarily written for coco dataset which contains human pose keypoints. So with few changes to it, we can use it for any custom dataset.
To do that, edit the _COCO_KEYPOINTS_NAMES list in line no 87 with our keypoints data with the same order it appears on the annotation file. This is an important step which needs to be verified.
Hello, @AmazingAbhishek
This solution is work for. If work so please select as solution for this suggestion.
Thanks,