Reduce inference time of object detection model by retraining with subset of original dataset

user_13956685 0
Viewed 413

#newb

I'm using the MobileNet SSD v2 (COCO) object detection model to detect trains from a live camera feed.

My goal is to reduce the inference time.

The COCO dataset is available here - there's 3745 annotated images of trains.

I'm using this tutorial by Coral to retrain the mobilenet model.

So if I create a dataset using the train images from COCO and retrain this specific model, will the inference time be reduced? Or is creating a new model from scratch the only way?

object-detection google-coral tensorflow computer-vision
1 Answers

Second @Shubham's answer, this won't make any difference in inference time, as long as you keep the input/output size the same. After the retraining the model, you'll have a fully quantized tflite model, as long as you follow the tutorial and compile it for the edgetpu, you can expects similar results to this benchmark.

user_62627570
Related