eval_semantic_segmentation is wrong about chainercv HOT 2 CLOSED

Since it would be too memory intensive to accept list of labels of the entire dataset as input, I am considering to split the function into two.
The first one takes list or batch of images and returns confusion matrix (i.e. hist).
The second one takes the matrix and calculate acc, acc_cls, miou, fwavacc.

I will also write a SemanticSegmentationEvaluator that does following in a loop

1. get batch_imgs and batch_gt_labels
2. pred_labels = model.predict(batch_imgs)
3. hist += compute_hist(pred_labels, batch_gt_labels)
4. (after the entire dataset is sampled), `pa = compute_pixel_accuracy(hist); ma = compute_mean_accuracy(hist); miou = compute_miou(hist) ...

from chainercv.

Fixed #217

from chainercv.