bbox_regressor.py 文件源码

def train_bbox_regressor(X, bbox, gt):
  config = Data()
  config.min_overlap = 0.6
  config.delta = 1000
  config.method = 'ridge_reg_chol'

  # get training groundtruth
  Y, O = get_examples(bbox, gt)
  X = X[O>config.min_overlap]
  Y = Y[O>config.min_overlap]

  # add bias
  X = np.c_[X, np.ones([X.shape[0], 1])]

  # center and decorrelate targets
  mu = np.mean(Y, axis=0).reshape(1, -1)
  Y = Y - mu
  S = dot(Y.T, Y) / Y.shape[0]
  D, V = eig(S)
  T = dot(dot(V, diag(1.0/sqrt(D+0.001))), V.T)
  T_inv = dot(dot(V, diag(sqrt(D+0.001))), V.T)
  Y = dot(Y, T)

  model = Data()
  model.mu = mu
  model.T = T
  model.T_inv = T_inv
  model.Beta = np.c_[solve(X, Y[:, 0], config.delta, config.method),
                     solve(X, Y[:, 1], config.delta, config.method),
                     solve(X, Y[:, 2], config.delta, config.method),
                     solve(X, Y[:, 3], config.delta, config.method)]

  # pack 
  bbox_reg = Data()
  bbox_reg.model = model
  bbox_reg.config = config
  return bbox_reg