Hi. Thank you for awesome Edge Impulse. Our company very interested in creating ML Model for edge device, especially Himax WE-1 board.
I generated 3 models which detected dog versus cat, real golden retriever versus doll golden retriever, and doll cat versus real cat.
- Dog versus Cat
-
dataset from Kaggle website(https://www.kaggle.com/tongpython/cat-and-dog?select=test_set)
-
Impulse Design : Image data - Image - Transfer Learning - Output feature. (MobileNetV2 0.35(16 neurons, 0.1 dropout) / Training cycle 10 / Learning rate 0.0005)
-
Model test data : cat(1010) dog(1012) uncertain(105)
uncertain data consist of fruit data, kite, horses, person .jpg. This data is unseen data.
= Expected result is classifying cat, dog, uncertain.
e.g
sample name | Expected outcome | Accuracy | Result
cherry | uncertain | 0% | 1 real
plums | uncertain | 0% | 1 real
The model test result is poor. uncertain data was not classified.
- real golden retriever versus doll golden retriever.
I want to check the performance of model can classify real golden retriever and doll golden retriever.
-
dataset captured by Himax WE-1 camera sensor.
-
Impulse Design : Image data - Image - Transfer Learning - Output feature. (MobileNetV2 0.35(16 neurons, 0.1 dropout) / Training cycle 10 / Learning rate 0.0005)
-
Model test : golden retriever doll, real golden retriever picture, other dog picture, cat picture, eagle, horses, scream
e.g
sample name | Expected outcome | Accuracy | Result
cat | uncertain | 0% | 1 real
horse | uncertain | 0% |1 real
uncertain data was not classified either.
- doll cat versus real cat
-
dataset : cat doll(captured by Himax WE-1 board camera sensor), real cat data(from kaggle)
-
Impulse Design : Image data - Image - Transfer Learning - Output feature. (MobileNetV2 0.35(16 neurons, 0.1 dropout) / Training cycle 10 / Learning rate 0.0005)
-
Model test : real cat picture, doll cat picture, other dog picture
Analysis of the three model tests is as follows: The doll and the real compare very well. For example, if you are training a dog doll and a real dog, it is recognized as a doll if you put a cat doll as an input. If you enter a cat, it will be recognized as a real dog. However, if you test with completely different data such as horse or kite, or data that requires an uncertain result value, the result is a real puppy.
I want to improve the model performance, is there any way?
For input data, about 8000 sheets were used when downloaded from kaggle, and about 200 sheets were used when taken directly using the himax device.
Please advice for me. Thanks