search_tools.py 文件源码

def reduce_distmat(full_dist_mat,
                   gal_templateids,
                   probe_templateids,
                   reduce_type=ReduceType.MeanMin):
  # Get unique template indices and there positions for keeping initial order
  #gal_tuids,gal_tuind=np.unique(gal_templateids,return_index=True)
  #probe_tuids,probe_tuind=np.unique(probe_templateids,return_index=True)
  gal_tuids, gal_tuind = np.unique(
      [str(x) for x in gal_templateids], return_index=True)
  probe_tuids, probe_tuind = np.unique(
      [str(x) for x in probe_templateids], return_index=True)
  red_dist_mat = np.zeros((len(gal_tuids), len(probe_tuids)))
  # Loop on gallery
  for g, gtupos in enumerate(gal_tuind):
    gutid = gal_templateids[gtupos]
    gt_pos = np.where(gal_templateids == gutid)[0]
    # Loop on probe
    for p, ptupos in enumerate(probe_tuind):
      putid = probe_templateids[ptupos]
      pt_pos = np.where(probe_templateids == putid)[0]
      # Get appropriate distance
      #print g,p
      dist_val = 0.0
      # TO BE FIXED
      if reduce_type == ReduceType.MeanMin:
        dist_val = np.mean(np.min(full_dist_mat[np.ix_(gt_pos, pt_pos)]))
      else:
        dist_val = np.amin(full_dist_mat[np.ix_(gt_pos, pt_pos)])
      red_dist_mat[g, p] = dist_val
  return red_dist_mat, gal_tuind, probe_tuind