After my previous blog post it was pointed out to me that the amount of whitespace (or other non-object pixels, i.e. background) that I had in my captured image was affecting the accuracy of the classification. I did a quick inverse test that was suggested by beacon_dave and added whitespace around the CIFAR-10 test image and it then classified as an airplane! I guess this makes sense as there are a lot of extraneous pixels to confuse the classifier. In general purpose use of a camera or other image source, the object to be classified would be normally identified and enclosed in a tight bounding box and only the pixels in the box would be used to classify the object. For this example notebook that step was left out. I think the assumption was that you could create that bounding box by zooming in on the object (which I not sure how to do with my webcam). So, I tried the next best thing and repeated the test using cropped images from the webcam. And that produced more satisfactory results. Using neural networks and training them correctly is indeed a challenging task.
CIFAR-10 Sportscar test image with added whitespace
Webcam Firetruck cropped image