diff --git a/rcnn/symbol/symbol_mask_fpn.py b/rcnn/symbol/symbol_mask_fpn.py
index a7cae67..da29e78 100755
--- a/rcnn/symbol/symbol_mask_fpn.py
+++ b/rcnn/symbol/symbol_mask_fpn.py
@@ -81,20 +81,23 @@ def get_resnet_conv_down(conv_feat):
     P4_la   = mx.symbol.Convolution(data=conv_feat[1], kernel=(1, 1), num_filter=256, name="P4_lateral")
     P5_clip = mx.symbol.Crop(*[P5_up, P4_la], name="P4_clip")
     P4      = mx.sym.ElementWiseSum(*[P5_clip, P4_la], name="P4_sum")
-    P4      = mx.symbol.Convolution(data=P4, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P4_aggregate")
 
     # P4 2x upsampling + C3 = P3
     P4_up   = mx.symbol.UpSampling(P4, scale=2, sample_type='nearest', workspace=512, name='P4_upsampling', num_args=1)
     P3_la   = mx.symbol.Convolution(data=conv_feat[2], kernel=(1, 1), num_filter=256, name="P3_lateral")
     P4_clip = mx.symbol.Crop(*[P4_up, P3_la], name="P3_clip")
     P3      = mx.sym.ElementWiseSum(*[P4_clip, P3_la], name="P3_sum")
-    P3      = mx.symbol.Convolution(data=P3, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P3_aggregate")
 
     # P3 2x upsampling + C2 = P2
     P3_up   = mx.symbol.UpSampling(P3, scale=2, sample_type='nearest', workspace=512, name='P3_upsampling', num_args=1)
     P2_la   = mx.symbol.Convolution(data=conv_feat[3], kernel=(1, 1), num_filter=256, name="P2_lateral")
     P3_clip = mx.symbol.Crop(*[P3_up, P2_la], name="P2_clip")
     P2      = mx.sym.ElementWiseSum(*[P3_clip, P2_la], name="P2_sum")
+
+    # P5,P4,P3,P2 aggregate: reduce the aliasing effect of upsampling
+    P5      = mx.symbol.Convolution(data=P5, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P5_aggregate")
+    P4      = mx.symbol.Convolution(data=P4, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P4_aggregate")
+    P3      = mx.symbol.Convolution(data=P3, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P3_aggregate")
     P2      = mx.symbol.Convolution(data=P2, kernel=(3, 3), pad=(1, 1), num_filter=256, name="P2_aggregate")
 
     # P6 2x subsampling P5