/tflite-object-detection

Tensorflow Lite object detection

Primary LanguagePython

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