Tensorflow lite object detection with SSD MobileNet v2

Inference

Inference requires the following steps

1. Prepare input:

   • resize to (300, 300) for SSD MobileNet
   • input normalization using mean and standard deviation
   • input tensor accepts NxHxWxC where n=1 for single image

2. Run Inference

3. Post process output:

   • SSD MobileNet returns 10 detection by default. The 4th output tensor indicates how many are valid.
   • For the valid detection, convert the boundary box coordinates to the scale of image size.

4. Evaluate using metrics

   • Use tensorflow metrics (models/research/object_detection/metrics) API to evaluate the detections