How to increase fps in tensorflow object detection model

TensorFlow General Discussion
, , ,
1

i have a trained tensorflow model , for the training i select ssdmobilenetfpnlite(640x640) after that i convert my model to tflite. while inferencing i got only 3-4 fps how i increase that fps more than 25

3

Hi @Amit_Kumar_Tiwaree ,
Here are a few possibilities that you can increase Frames per second(fps) while inferencing:
From TFLIte point of view:

  1. Optimize the TFLite model to int8 quantization(improves the speed by 3x+) Ref.
  2. If your target device has hardware acceleration, use it to run your model. The supported delegates for hardware acceleration are: NNAPI on Android, GPU delegates on iOS and Android, and Edge TPU on Coral devices.
    From Base model point of view:
  3. Use the model optimization techniques like Pruning and knowledge distillation.
  4. If acceptable to the accuracy barometer, try to reduce the input image resolution to 320x320. It will significantly improve fps especially on low-end devices.

Thank You