Inference requires the following steps
-
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
-
Run Inference
-
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.
-
Evaluate using metrics
- Use tensorflow metrics (models/research/object_detection/metrics) API to evaluate the detections