Training Custom Object Detection Model with TensorFlow Object Detection API

Hi community! I have tried to train my first custom object detection model with faster_rcnn_inception_resnet_v2. I have 10 classes, and my dataset consists 50 images for each class. Each image is annotated with LabelImg.

However, I have a few questions here:

  • The initial total loss is low (starts from 1.1). I eventually stopped the training process at 5000 steps as the total loss has already dropped to below 0.1.

  • But, a lot of objects are not detected and drawn with bounding boxes as the end result.

  • And, some detected objects consist of very low confidence score. (Below 50%)

May I ask if there is any ways that I can further improve my model?

Could you try to increase the number of images in your training dataset? Thank you

My dataset consists of handdrawn images which are drawn manually. The annotations are also done manually.

In the pipeline.config file, data augmentation techniques are implemented. I’m just wondering if these will help to generate more images in the dataset.

PS: I am trying to work on a project who does the same as the link below: