avg_stats, all_extra_info: To undertand the content of these structs, we
 should know that RIGOR performs parametric min-cut over different graph 
 types, each of which specifies unaries differently. The two major graph 
 types are uniform and color graphs (see code). Uniform graphs define 
 unaries based on the size of the superpixel. Color graphs define unaries 
 based on the color difference to the seed superpixels. Each major graph 
 type has multiple sub-types - each defining unaries slightly differently. 
 RIGOR has 3 different types of uniform graphs and 3 different color 
 graphs. In total RIGOR (CVPR'14) has G=6 graph-types.


 <all_extra_info>: stores quantitative scores about all N ground-truths 
  |            (over I images):
  |-- <im_no> is a 1xN vector giving the image index for each GT.
  |-- <gt_category_id> is a 1xN vector which gives the category id for 
  |            each GT.
  |-- <best_overlaps> is a 1xN vector of best overlaps for each GT. The 
  |            best overlapping segment is selected from the set of 
  |            proposals generated for image im_no(i).
  |-- <best_ov_seg_idx> is a 1xN vector giving the best overlap proposal 
  |            index for each GT i.e. this proposal, for image <im_no>(i), 
  |            gives the best overlap score for the current GT.
  |-- <best_overlaps_sgraphs> is a NxG matrix, where each column gives 
  |            best overlaps for one graph sub-type. 
  |            max(best_overlaps_sgraphs,[],2) == best_overlaps'
  |-- <sgraph_gt_wins> is a IxG matrix, where each column gives the number 
  |            of times a graph sub-type gives the best overlap (the 
  |            winning proposal). sum(sgraph_gt_wins,2) is the number of 
  |            GTs in each image.
  |-- <gt_szs> is a 1xN vector giving the size in pixels for each GT.
  |-- <bb_overlaps> is similar to <best_overlaps>, except that it is over 
  |            bounding boxes (of a GT segment) rather than the GT 
  |            segmentation.
  |-- <best_bbov_seg_idx> is the bounding box counterpart of 
  |            <best_ov_seg_idx>.
  |-- <bb_szs> is the bounding box counterpart of <gt_szs> (size of the 
  |            bounding box rectangle).
  |-- <sgraph_names> is a 1xG cell array, where each cell is a string 
  |            identifying a graph sub-type.


 <avg_stats>: stores average quantitative scores and some other 
  |            information:
  |-- <final_num_segs> Gives the average number of proposals generated 
  |            over I images in the dataset.
  |-- <avg_best_overlap> Is the average overlap over all GTs. 
  |            = mean(all_extra_info.best_overlaps)
  |-- <covering> Is the covering computed by averaging the covering scores 
  |            for all I images.
  |-- <collated_covering> Covering computed by considering all GTs 
  |            together i.e.
  |            = sum(all_extra_info.best_overlaps.*all_extra_info.gt_szs) 
  |              / sum(all_extra_info.gt_szs);
  |-- <avg_best_overlaps_sgraphs> Is a 1xG vector average overlap scores 
  |            over each graph sub-type, i.e. 
  |            = mean(all_extra_info.best_overlaps_sgraphs)
  |-- <all_sgraph_gt_wins> Is a 1xG vector giving the number of times each 
  |            graph sub-type produces a proposal with overall best 
  |            overlap, i.e. = sum(all_extra_info.sgraph_gt_wins). 
  |            Also sum(all_sgraph_gt_wins) = N
  |-- <avg_bb_overlap> The bounding box counterpart of <avg_best_overlap>, 
  |            i.e. = mean(all_extra_info.bb_overlaps)
  |-- <num_gts> = N
  |-- <num_ims> = I
  |-- <gt_categories> Is a 1xC cell array, which gives names of each of 
  |            the C GT categories. The ordering is according to the ids 
  |            in <gt_category_id>
  |-- <category_avg_overlaps: Is a 1xC array, which gives the average 
  |            best overlap for each GT category i.e. 
  |            = accumarray(all_extra_info.gt_category_id(:), 
  |                         all_extra_info.best_overlaps, [], @mean)
  |-- <num_category_gts> Is the number of instances for each GT category, 
  |            i.e. = accumarray(all_extra_info.gt_category_id(:), 1)
  |-- <seg_recall> Is a 1x3 array giving recall/detection rates at 0.5, 
  |            0.7, 0.9 overlaps, i.e. percentage of GTs which have an 
  |            overlap above 0.5, 0.7, or 0.9. For instance at 0.5 the 
  |            recall is computed as 
  |            nnz(all_extra_info.best_overlaps>=0.5) / avg_stats.num_gts
  |-- <bb_recall> Is the bounding box counterpart of <seg_recall>.