4 ms·
Those were some interesting failure cases, but the conclusion is unwarranted. These mistakes are not because of any shortcoming of Inception models (except mayb
by ma2rten 6y ago
Those were some interesting failure cases, but the conclusion is unwarranted. These mistakes are not because of any shortcoming of Inception models (except maybe the upside-down car example). They are because the examples are out of domain for the training data, the model just has never seen sliced apples. Commercial providers likely have their own private training data.
BTW, it would be interesting to see how OpenAI's CLIP does on this problem, since it has more varied training data.
- mbil 6y agoYeah the car example also struck me as a real shortcoming -- sure, the training set probably included very few if any images of upside down cars, but I would expect basic data augmentation to have addressed that.
- bonoboTP 6y ago180 degree rotation is usually not included in augmentation because it hurts performance in upright images. You can do it, but if you want high benchmark scores it's better to limit rotation augmentation to smaller amounts, usually less than 45 degrees.