Problem with choosing bacis model to train custom model to object detecion on mobile devices (tf lite))

Hi everyone.

I’m a mobile app developer.

I need to prepare a model that searches for objects in a photo (on mobile devices).
I want to train the model on my own data. The final format is tensorflow lite.

I have not used tensorflow lite or any other technology to train/recognize images at all before.
I have a few questions:

  1. what initial model would you recommend? I have read about SSD and YOLO.
  2. would you recommend some resources to better understand the issues with training and image recognition?
  3. please evaluate this post Object Detection with TensorFlow Lite Model Maker ? Is EfficientDet better or worse than the SSD model? Objects that my target model after training is to recognize are climbing holds (different shapes, sizes). Can I use EfficientDet?

Hi @Chris55521,

  • Both SSD and YOLO are good choices for object detection and where high speed and accuracy are considered. SSD is generally a good choice for object detection on Android devices.
  • Basically TFLite is designed for inferencing on edge devices. Once the model is trained for object detection using either SSD or YOLO, then the model is converted to TFLite for inferencing. The TFLitr file is light weight file but equally outputs the accuracy as like the model. Please go through the links for additional information link1 and examples link2.
  • For recognizing climbing holds of different shapes and sizes in your use case, you can use EfficientDet for high accuracy on android devices but little bit slower than SSD and YOLO.
  • At the starting level you can proceed with TFLite model maker for object detection. Model Maker library simplifies the process of training a TensorflowLite model using custom dataset. Go through the ref. Try model maker with Python <3.10.

Thank You