diff --git a/PaddleCV/object_detection/ppdet/modeling/backbones/resnet.py b/PaddleCV/object_detection/ppdet/modeling/backbones/resnet.py
index c902b8acd2ee8d4729066add4d0ab88e3d0b25b2..7d879ba7521370fa3ae2eda759da1efe850b713d 100644
--- a/PaddleCV/object_detection/ppdet/modeling/backbones/resnet.py
+++ b/PaddleCV/object_detection/ppdet/modeling/backbones/resnet.py
@@ -152,8 +152,7 @@ class ResNet(object):
         ch_in = input.shape[1]
         # the naming rule is same as pretrained weight
         name = self.na.fix_shortcut_name(name)
-
-        if ch_in != ch_out or stride != 1:
+        if ch_in != ch_out or stride != 1 or (self.depth < 50 and is_first):
             if max_pooling_in_short_cut and not is_first:
                 input = fluid.layers.pool2d(
                     input=input,
@@ -252,6 +251,8 @@ class ResNet(object):
         conv = input
         for i in range(count):
             conv_name = self.na.fix_layer_warp_name(stage_num, count, i)
+            if self.depth < 50:
+                is_first = True if i == 0 and stage_num == 2 else False
             conv = block_func(
                 input=conv,
                 num_filters=ch_out,