losses collection (step, bbox, iou, class, total) will be saved in logs/losses_collection.txt
edit adamop to use AdamOptimizer instead of SGD with momentum (default) and pretrained to use VGG16 model from tf-slim instead of initialization from scratch
validation - testing:
demo:
todo:
multiscale images training
fixed out of GPU's memory while training, exhausted with tf.ConfigProto GPU options
evaluate model and visualization with matplotlib
using compute targets (groundtruth and mask) with cython to improve training speed