TODO

train models with natural aspect ratio
cleanup and update debugging notebooks
detection and recognition in one model?

change naming
    true_labels, pred_labels
    input/image size?
    'plot' in 'draw'
    'inter layer segments'

data augmentation
    augmentation and batch size in validation?
    add random rotation, translation, zero padding

SSD
    check if prior boxes are always at the same location in the receptive field
        is it related to bad results on small objects?
    convert fc6, fc7 vgg layers to conv layers
        training time is much longer, is it caused by the missing pretrained fc layers?
    more native image aspect ratio 128 * (4, 3) = (512, 384), but largest box is not in center
    ResNet https://github.com/hzm8341/SSD_keras_restnet/blob/master/SSD_ResNet/SSD_resnet.py
    MobileNet https://github.com/tanakataiki/ssd_keras/blob/master/ssd300MobileNet.py
    do not crop boxes outside image
    remove flip option for aspect ratios
    try input normalisation
    unfreeze layers gradually?
    DONE
        use softmax for each class, not categorical cross entropy
        DSOD
        precision, recall, f-mean / f-measure
        GT background class
        sample images from generator
        https://github.com/scardine/image_size
        no one_hot in gt_util
        CustomCallbacks in ssd_traing, for learning rate and logging
        try focal loss, with alpha = 1 and gamma=3 can boost 1.4% mAP on pascal voc 2007
        custom epoch size, checkpoint callback, test, eval interval? 
        improve decoding performance, sparse, faster nms
        hard negative mining?, fix it
        virtual batch_size

SegLink
    width of segment ground truth?
    
    directed links to get text orientation
    training data, on icdar combined words to lines lead to better results
    
    display overlap of segments
    
    training
        lambdas = 1
        first train segments, then linking
        parameters for sl loss and focal loss
    
    debug
        segment ground truth at the left and right side
        check if parameters get properly loaded
        why some boxes are wrong
    
    prior box is assigned to ground truth with smallest distance (if it lies in multiple gt boxes), is this the case?
    
    DONE
        grid search
        confidence of combined boxes is the mean conf of the segments weighted by there area
        more natural input ratio

TextBoxes / TextBoxes++
    use recognition score to reduce false positives in detection

CRNN
    larger / variable input width
    more units
    RGB input?
    check synthtext? 'small caps' etc
    normalize input, subtract mean
    DONE
        larger alphabet
        GRU
        editdistance / levenshtein distance