This is mockup pipeline for an image classification task that will reside on an NVIDIA Jetson device

This pipeline proceeds by sequentially processing each JPG/PNG image in four primary steps:

1. YOLOv5 object detection algorithm to identify motor vehicles (i.e. a car, truck, or bus) in an image
   • Generate bounding box coordinates for all vehicles
   • Restrict to the largest vehicle (in square pixels) if multiple vehicles are detected
   • Dilate the bounding box by 5 pixels, since the bounding boxes output by YOLOv5 tend to be very tight
     
     
2. Crop and resize the image array
   • The image array is cropped to the dilated bounding box coordinates
   • Resize the cropped image array to 224 x 224 pixels to adhere to the dimension of the pretrained ResNet50 model in step #3
   • This portion of the pipeline is done using a TensorFlow dataset, though the image is converted back to a numpy array for step #3
     
     
3. Feed the image array to the vehicle make-model classifier
   • This model was developed and trained as a TensorFlow Keras model
   • The final model weights were then converted to Onnx format
   • This model outputs a numpy array of softmax probabilities equal to the number of classes (n=574). Taking the argmax, we obtain the predicted vehicle make-model class
     
     
4. Affix the bounding box and make-model label to the original image array and output as a PNG
   • Examples: