dont use grad_x for pixels selected for grad_y and vice-versa: 3x fewer alignment failures

Author: catid

alignment.cpp

@@ -678,7 +678,7 @@ bool VideoAligner::AlignNextFrame(
 #endif
 
         if (i > 0) {
-            // Move from half-resolution to full-resolution
+            // Move up the pyramid to 2x higher resolution
             transform.TX *= 2.0;
             transform.TY *= 2.0;
         }

generators.cpp

@@ -350,18 +350,20 @@ class sparse_jac_generator : public Halide::Generator<sparse_jac_generator>
 
         Expr scale = 1.f / grad_x.width();
 
+        // Sparse pixels selected based on grad_x should only use grad_x in calculation
         output_x(x, y, c) = mux(c, {
-            (grad_x(ix0, iy0) * ix0 + grad_y(ix0, iy0) * iy0) * scale,
-            (grad_x(ix0, iy0) * -iy0 + grad_y(ix0, iy0) * ix0) * scale,
-            (grad_x(ix0, iy0)),
-            (grad_y(ix0, iy0)),
+            2.f * grad_x(ix0, iy0) * ix0 * scale,
+            2.f * grad_x(ix0, iy0) * -iy0 * scale,
+            2.f * grad_x(ix0, iy0),
+            0.f,
         });
 
+        // Sparse pixels selected based on grad_y should only use grad_y in calculation
         output_y(x, y, c) = mux(c, {
-            (grad_x(ix1, iy1) * ix1 + grad_y(ix1, iy1) * iy1) * scale,
-            (grad_x(ix1, iy1) * -iy1 + grad_y(ix1, iy1) * ix1) * scale,
-            (grad_x(ix1, iy1)),
-            (grad_y(ix1, iy1)),
+            2.f * grad_y(ix1, iy1) * iy1 * scale,
+            2.f * grad_y(ix1, iy1) * ix1 * scale,
+            0.f,
+            2.f * grad_y(ix1, iy1),
         });
     }
 

schedules/sparse_jac.schedule.h

@@ -19,7 +19,7 @@ inline void apply_schedule_sparse_jac(
     using ::Halide::TailStrategy;
     using ::Halide::Var;
     Func output_y = pipeline.get_func(5);
-    Func output_x = pipeline.get_func(3);
+    Func output_x = pipeline.get_func(2);
     Var v0(output_y.get_schedule().dims()[0].var);
     Var v0i("v0i");
     Var v1(output_y.get_schedule().dims()[1].var);