Training using multiple datasets

[VittalP]

This branch provides support to training Faster RCNN using multiple datasets; e.g. voc_2007_trainval+voc_2012_trainval (07+12). This is achieved by replacing the get_roidb function with a combined_roidb which returns one or more roidbs based on the input imdb name.

Below are the results on voc_2007_test when trained on "07+12" as described in the Faster RCNN paper.

VOC07 metric? Yes
AP for aeroplane = 0.7612
AP for bicycle = 0.7916
AP for bird = 0.7116
AP for boat = 0.6001
AP for bottle = 0.5101
AP for bus = 0.8355
AP for car = 0.8333
AP for cat = 0.8883
AP for chair = 0.4977
AP for cow = 0.7875
AP for diningtable = 0.6350
AP for dog = 0.8409
AP for horse = 0.8381
AP for motorbike = 0.7674
AP for person = 0.7565
AP for pottedplant = 0.3792
AP for sheep = 0.7595
AP for sofa = 0.6932
AP for train = 0.8040
AP for tvmonitor = 0.6781
Mean AP = 0.7184

Results:
0.761
0.792
0.712
0.600
0.510
0.836
0.833
0.888
0.498
0.787
0.635
0.841
0.838
0.767
0.756
0.379
0.759
0.693
0.804
0.678
0.718

The results are 1.4% below the numbers quoted in the ArXiv paper ( http://arxiv.org/pdf/1506.01497v3.pdf ).