Relax the requirements on loss functiond derivatives.
We now require that the first derivative of the loss function
be positive only if the second derivative is non-zero. This is
because when the second derivative is non-positive, we do not use
the second order correction suggested by BANS and instead use
a simpler first order strategy which does not use a division by
the gradient of the loss function.
Change-Id: I3d65713f152611998e196ff389a7081acfdfd8c1
diff --git a/internal/ceres/corrector.cc b/internal/ceres/corrector.cc
index 955feb5..cbf49ec 100644
--- a/internal/ceres/corrector.cc
+++ b/internal/ceres/corrector.cc
@@ -40,7 +40,6 @@
Corrector::Corrector(const double sq_norm, const double rho[3]) {
CHECK_GE(sq_norm, 0.0);
- CHECK_GT(rho[1], 0.0);
sqrt_rho1_ = sqrt(rho[1]);
// If sq_norm = 0.0, the correction becomes trivial, the residual
@@ -85,6 +84,14 @@
return;
}
+ // We now require that the first derivative of the loss function be
+ // positive only if the second derivative is non-zero. This is
+ // because when the second derivative is non-positive, we do not use
+ // the second order correction suggested by BANS and instead use a
+ // simpler first order strategy which does not use a division by the
+ // gradient of the loss function.
+ CHECK_GT(rho[1], 0.0);
+
// Calculate the smaller of the two solutions to the equation
//
// 0.5 * alpha^2 - alpha - rho'' / rho' * z'z = 0.
