diff --git a/src/shogun/metric/LMNN.cpp b/src/shogun/metric/LMNN.cpp
index 9fddd7274f8..95546f035f2 100644
--- a/src/shogun/metric/LMNN.cpp
+++ b/src/shogun/metric/LMNN.cpp
@@ -97,8 +97,7 @@ void CLMNN::train(SGMatrix<float64_t> init_transform)
 
 		// Compute objective
 		SG_DEBUG("Computing objective.\n")
-		obj[iter] = CLMNNImpl::compute_objective(x, L, outer_products, target_nn,
-				cur_impostors, m_regularization);
+		obj[iter] = CLMNNImpl::compute_objective(L, gradient);
 
 		// Correct step size
 		CLMNNImpl::correct_stepsize(stepsize, obj, iter);
diff --git a/src/shogun/metric/LMNNImpl.cpp b/src/shogun/metric/LMNNImpl.cpp
index 8881ecea5db..72222b5b47a 100644
--- a/src/shogun/metric/LMNNImpl.cpp
+++ b/src/shogun/metric/LMNNImpl.cpp
@@ -245,36 +245,12 @@ void CLMNNImpl::gradient_step(MatrixXd& L, const MatrixXd& G, float64_t stepsize
 	L -= stepsize*(2*L*G);
 }
 
-float64_t CLMNNImpl::compute_objective(const CDenseFeatures<float64_t>* x, const MatrixXd& L,
-		const OuterProductsMatrixType& C, const SGMatrix<index_t> target_nn,
-		const ImpostorsSetType& Nc, float64_t regularization)
+float64_t CLMNNImpl::compute_objective(const MatrixXd& L, const MatrixXd& G)
 {
-	// get the number of examples from data
-	int32_t n = x->get_num_vectors();
-	// get the number of target neighbors per example (k) from the arguments
-	int32_t k = target_nn.num_rows;
-	// initialize the objective
-	float64_t obj = 0;
 	// pre-compute the Mahalanobis distance matrix
 	MatrixXd M = L.transpose()*L;
-
-	// add pull contributions to the objective
-	for (int32_t i = 0; i < n; ++i) // for each training example
-	{
-		for (int32_t j = 0; j < k; ++j) // for each target neighbor
-			obj += (1-regularization)*TRACE(M,C[i][ target_nn(j,i) ]);
-	}
-
-	// add push contributions to the objective
-	for (ImpostorsSetType::iterator it = Nc.begin(); it != Nc.end(); ++it) // for each possible impostor
-	{
-		double hinge = 1 + TRACE(M,C[it->example][it->target]) - TRACE(M,C[it->example][it->impostor]);
-
-		if (hinge > 0)
-			obj += regularization*hinge;
-	}
-
-	return obj;
+	// compute objective
+	return TRACE(M,G);
 }
 
 void CLMNNImpl::correct_stepsize(float64_t& stepsize, const SGVector<float64_t> obj, const uint32_t iter)
diff --git a/src/shogun/metric/LMNNImpl.h b/src/shogun/metric/LMNNImpl.h
index 2332894b8b7..26e487ba335 100644
--- a/src/shogun/metric/LMNNImpl.h
+++ b/src/shogun/metric/LMNNImpl.h
@@ -109,7 +109,7 @@ class CLMNNImpl
 		static void gradient_step(Eigen::MatrixXd& L, const Eigen::MatrixXd& G, float64_t stepsize);
 
 		/** compute LMNN objective */
-		static float64_t compute_objective(const CDenseFeatures<float64_t>* x, const Eigen::MatrixXd& L, const OuterProductsMatrixType& C, const SGMatrix<index_t> target_nn, const ImpostorsSetType& Nc, float64_t mu);
+		static float64_t compute_objective(const Eigen::MatrixXd& G, const Eigen::MatrixXd& L);
 
 		/** correct step size depending on the last fluctuation of the objective */
 		static void correct_stepsize(float64_t& stepsize, const SGVector<float64_t> obj, const uint32_t iter);
diff --git a/tests/unit/metric/LMNN_unittest.cc b/tests/unit/metric/LMNN_unittest.cc
index 386b7edc902..3973db7a9b8 100644
--- a/tests/unit/metric/LMNN_unittest.cc
+++ b/tests/unit/metric/LMNN_unittest.cc
@@ -56,7 +56,7 @@ TEST(LMNN,train)
 
 	// check linear transform solution
 	SGMatrix<float64_t> L=lmnn->get_linear_transform();
-	EXPECT_NEAR(L(0,0),0.983061830081943477,1e-5);
+	EXPECT_NEAR(L(0,0),0.991577280560543,1e-5);
 	EXPECT_NEAR(L(0,1),0,1e-5);
 	EXPECT_NEAR(L(1,0),0,1e-5);
 	EXPECT_NEAR(L(1,1),1.00000080000000002,1e-5);