Object recognition at different angles

This was the very first time I played around with an image classifier, just a week after starting my Bachelor’s degree in Artificial intelligence and Data. Back then I had no idea how the model actually worked, but accepted is as a black box.

Basically the experiment was to use an out-of-the box Resnet50 model to recognise a phone at different angles. Another purpose of the experiment was to get familiar with means and confidence intervals.

This was gathered in a table

Angle (deg.)							Avg. (%)
90	y	y	y	y	y	y	100
80	y	y	y	y	y	y	100
70	y	y	y	y	y	y	100
60	y	y	y	y	y	y	100
50	y	y	y	y	y	y	100
40	y	y	y	y	y	y	100
30	y	y	y	y	y	y	100
20	y	y	y	y	y	y	100
10	n	n	n	y	y	n	33.33
0	n	y	n	n	n	n	16.7

We collected these results and found an average accuracy of 85% with a confidence interval [80.6%; 89.4%].

All this was as simple as the following code and some boilerplate to load the image

1
2
3
4
5
6
model = resnet50.ResNet50(weights='imagenet')
# load the image
# ....
processed_image = resnet50.preprocess_input(np.expand_dims(frame, axis=0))
predictions = model.predict(processed_image)
label = resnet50.decode_predictions(predictions)    