We have tried using logistic regression (LR) to segment the images.
This produces the results shown in the following 2 figures.
While LR performed comparably with DRF in one context, it is significantly worse in the other.
The above figure shows results of LMU active
learning algorithm for geometric context dataset, using the LR segmentor.
Here, the performance of the LR segmentor is comparable to DRFs.
The above figure shows the results of LMU
active learning algorithm for brain tumor dataset, using the LR segmentor.
Here, DRF significantly outperforms LR... probably because the labels for
each pixel in the brain image is very dependent on the labels of its neighboring
pixels.
In above figure, black line shows the results of LMU that uses logistic regression to select the first image to give to the oracle.
(This system then trains the DRF model, using that image.)
Similar to LMU with DRF, this active learning algorithm also reaches the
"all" line in 2 iterations.
Although its performance is not a good as LMU, it still outperforms the random approach.
The results of LMU algorithm that starts with a random selection instead of LU in its first step on brain dataset is shown in the following figure
As shown in the above figure, starting with a random image, particularly an image that may not contain enough information for training, would delay the learning process.
Although after few iterations, LMU will catch up the "all" line, as it did in this case after having 3 images in its training set.