diff --git a/configs/yolov3_mobilenet_v3.yml b/configs/yolov3_mobilenet_v3.yml
index 4463d6d27217a4d3567e73acbd9f7e58e6fcfbf2..223d14c4909998b6d76f54ff3b45d6d1badd9d1e 100644
--- a/configs/yolov3_mobilenet_v3.yml
+++ b/configs/yolov3_mobilenet_v3.yml
@@ -19,6 +19,7 @@ MobileNetV3:
   norm_decay: 0.
   model_name: large
   scale: 1.
+  extra_block_filters: []
   feature_maps: [1, 2, 3, 4, 6]
 
 YOLOv3Head:
diff --git a/docs/mobile_side/README.md b/docs/mobile_side/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..5db74476887d9036940acba2907ba917f51a6c1c
--- /dev/null
+++ b/docs/mobile_side/README.md
@@ -0,0 +1,33 @@
+# Practical Mobile-side detection method base on PaddleDetection
+
+Mobile-side models are provided base on following architecture:
+
+1. YOLOv3
+2. Cascade Faster RCNN
+3. SSD
+
+## YOLOv3 mobile-side model
+
+Mobile-side model based on YOLOv3 is a pruned model of YOLOv3-MobileNetv3, we pruned the YOLO-head of YOLOv3-MobileNetv3 and distill the pruned model by YOLOv3-ResNet34, which has a higher mAP on COCO as 31.4(input shape as 320\*320). For pruning, configurations as as follows:
+
+1. pruning YOLO-head with following configuration and the FLOPS pruned ratios is 86%.
+
+```
+--pruned_params="yolo_block.0.0.0.conv.weights,yolo_block.0.0.1.conv.weights,yolo_block.0.1.0.conv.weights,yolo_block.0.1.1.conv.weights,yolo_block.0.2.conv.weights,yolo_block.0.tip.conv.weights,yolo_block.1.0.0.conv.weights,yolo_block.1.0.1.conv.weights,yolo_block.1.1.0.conv.weights,yolo_block.1.1.1.conv.weights,yolo_block.1.2.conv.weights,yolo_block.1.tip.conv.weights,yolo_block.2.0.0.conv.weights,yolo_block.2.0.1.conv.weights,yolo_block.2.1.0.conv.weights,yolo_block.2.1.1.conv.weights,yolo_block.2.2.conv.weights,yolo_block.2.tip.conv.weights" \
+--pruned_ratios="0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.875,0.875,0.875,0.875,0.875,0.875"
+```
+
+2. pruning YOLO-head by [Filter Pruning via Geometric Median](https://arxiv.org/abs/1811.00250), use FPGM algorithm by setting:
+
+```
+--prune_criterion=geometry_median
+```
+
+### Model Zoo
+
+| Backbone         | prune method |     GFLOPs    | Model size(MB) | input shape |     teacher model     |   Box AP   | SD845 latency |                      download                          |
+| :----------------| :----------: | :-----------: | :------------: | :---------: | :-------------------: | :--------: | :-----------: |:-----------------------------------------------------: |
+| MobileNetv3      |   baseline   | 4.93          | 90.23          |     320     |           -           | 27.1       |     319ms     | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v3.pdparams) |
+| MobileNetv3      |    prune     | 0.66(-86.57%) | 16.21(-82.03%) |     320     | YOLOv3-ResNet34(31.4) | 24.6(-2.5) |      91ms     | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v3_mobilenetv3_prune75875_FPGM_distillby_r34.pdparams) |
+
+**NOTE：** `baseline` is the YOLOv3-MobileNetv3 base model, `prune` is the pruned model from YOLOv3-MobileNetv3 and pruned by configurations above, `Box AP` is test by `320*320` as input shape in both two models, and `latency` is test on Snapdragon845 with single thread. The pruned model is 2.5 times faster than the base model when the `Box AP` only decreased by 2.5.
diff --git a/slim/extensions/distill_pruned_model/distill_pruned_model.py b/slim/extensions/distill_pruned_model/distill_pruned_model.py
index 6ab641480604e29e0b32b000b49d15d3bb311f46..e5fa524a36b72f360326e0829875a09ee98a5b0b 100644
--- a/slim/extensions/distill_pruned_model/distill_pruned_model.py
+++ b/slim/extensions/distill_pruned_model/distill_pruned_model.py
@@ -231,7 +231,9 @@ def main():
     assert pruned_ratios > [0] * len(pruned_ratios) and pruned_ratios < [1] * len(pruned_ratios), \
         "The elements of pruned ratios should be in range (0, 1)."
 
-    pruner = Pruner()
+    assert FLAGS.prune_criterion in ['l1_norm', 'geometry_median'], \
+            "unsupported prune criterion {}".format(FLAGS.prune_criterion)
+    pruner = Pruner(criterion=FLAGS.prune_criterion)
     distill_prog = pruner.prune(
         fluid.default_main_program(),
         fluid.global_scope(),
@@ -361,5 +363,11 @@ if __name__ == '__main__':
         type=str,
         help="The ratios pruned iteratively for each parameter when calculating sensitivities."
     )
+    parser.add_argument(
+        "--prune_criterion",
+        default='l1_norm',
+        type=str,
+        help="criterion function type for channels sorting in pruning, can be set " \
+             "as 'l1_norm' or 'geometry_median' currently, default 'l1_norm'")
     FLAGS = parser.parse_args()
     main()
diff --git a/slim/prune/prune.py b/slim/prune/prune.py
index 606b29e7cc76ac6538737acde2fce2bb99bfd241..2dc6ea720af7a90921fdfefe9c7f90d5aee83633 100644
--- a/slim/prune/prune.py
+++ b/slim/prune/prune.py
@@ -187,7 +187,9 @@ def main():
             pruned_ratios < [1] * len(pruned_ratios)
             ), "The elements of pruned ratios should be in range (0, 1)."
 
-    pruner = Pruner()
+    assert FLAGS.prune_criterion in ['l1_norm', 'geometry_median'], \
+            "unsupported prune criterion {}".format(FLAGS.prune_criterion)
+    pruner = Pruner(criterion=FLAGS.prune_criterion)
     train_prog = pruner.prune(
         train_prog,
         fluid.global_scope(),
@@ -388,5 +390,11 @@ if __name__ == '__main__':
         default=False,
         action='store_true',
         help="Whether to only print the parameters' names and shapes.")
+    parser.add_argument(
+        "--prune_criterion",
+        default='l1_norm',
+        type=str,
+        help="criterion function type for channels sorting in pruning, can be set " \
+             "as 'l1_norm' or 'geometry_median' currently, default 'l1_norm'")
     FLAGS = parser.parse_args()
     main()