It is common to look at 90th or 99th percentile latency in computer systems.
A user won't notice the difference between a couple of milliseconds of lag but if a function occasionally takes several seconds that is very noticeable.
As a follow-up to @Cameron Chandler; if your RSS=0 it should trigger some alerts since it is 100% overfitting, thus evaluating that model does not make any sense as such. But yes, take the limit of $\log(x), x\rightarrow 0^+$ and notice it goes towards $-\infty$ thus you cannot find a more perfekt model - if you only look at AIC, but you can (almost) not ...
RSS = 0 implies that the model is a perfect fit since there was no residual. The limit of the log of 0 is $-\infty$, and since lower AICs are better, and this model is perfect,it makes perfect sense that the AIC should be a negative number such that no number can be lower.
Accuracy: The sum of the numbers on the diagonal divided by the sum of all numbers on the grid
Recall and Precision depend on if you want to take the micro or macro approach. See this blog post for more details (it gives a very similar example to your case): https://towardsdatascience.com/confusion-matrix-for-your-multi-class-machine-learning-model-...
Which value should I take into account for saying that my model has accuracy of ...?
None. Accuracy is practically meaningless in such class imbalanced settings; the metrics of interest here are precision, recall, and f1 score.
Now, it's true that the values of these metrics also fluctuate between runs, much similarly with the reported accuracy. But this is ...
Your calculation is correct, but you forgot to ask yourself one question: why should we consider "red" as the positive class? Precision and recall can be calculated for every class (i.e. considering the current class as positive), as opposed to accuracy.
So if we take "blue" as positive we get:
precision = NaN (because there's 0 ...
In my experience, the two approaches can give quite different results. It doesn't really show in the example you provided because the sizes are similar. However, in some cases in object detection, you can have the same object appear with very different sizes in two images. Your first approach will weigh the IoUs equally but the second will give more weight ...
Here are few things you can try to reduce overfitting:
Use batch normalization
add dropout layers
Increase the dataset
Use batch size as large as possible (I think you are using 32 go with 64)
to generate image dataset use flow from data
Use l1 and l2 regularizes in conv layers
If dataset is big increase the layers in neural network.
USE callbacks tf.keras....