diff --git a/README.md b/README.md
index 1e7687f1767e23bbf60535eecd5c215c39a97dcc..59dc84da4c6d6a5a6ac24d8b7dbaa0d338075b5e 100644
--- a/README.md
+++ b/README.md
@@ -74,4 +74,12 @@ The difference between this repo and gdrn conference version mainly including:
 `./core/gdrn_modeling/test_gdrn.sh <config_path> <gpu_ids> <ckpt_path> (other args)`
 
 ## Pose Refinement
-TODO: rudy
\ No newline at end of file
+We utilize depth information to further refine the estimated pose.
+We provide two types of refinement: fast refinement and iterative refinement.
+
+For fast refinement, we compare the rendered object depth and the observed depth to refine translation.
+Run
+
+`./core/gdrn_modeling/test_gdrn_depth_refine.sh <config_path> <gpu_ids> <ckpt_path> (other args)`
+
+For iterative refinement, please checkout to the pose_refine branch for details.
diff --git a/core/gdrn_modeling/test_gdrn_depth_refine.sh b/core/gdrn_modeling/test_gdrn_depth_refine.sh
new file mode 100644
index 0000000000000000000000000000000000000000..e242c6b4c4c49fd4bc1062d010d0eda6abb21ca0
--- /dev/null
+++ b/core/gdrn_modeling/test_gdrn_depth_refine.sh
@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# test
+set -x
+this_dir=$(dirname "$0")
+# commonly used opts:
+
+# MODEL.WEIGHTS: resume or pretrained, or test checkpoint
+CFG=$1
+CUDA_VISIBLE_DEVICES=$2
+IFS=',' read -ra GPUS <<< "$CUDA_VISIBLE_DEVICES"
+# GPUS=($(echo "$CUDA_VISIBLE_DEVICES" | tr ',' '\n'))
+NGPU=${#GPUS[@]}  # echo "${GPUS[0]}"
+echo "use gpu ids: $CUDA_VISIBLE_DEVICES num gpus: $NGPU"
+CKPT=$3
+if [ ! -f "$CKPT" ]; then
+    echo "$CKPT does not exist."
+    exit 1
+fi
+NCCL_DEBUG=INFO
+OMP_NUM_THREADS=1
+MKL_NUM_THREADS=1
+PYTHONPATH="$this_dir/../..":$PYTHONPATH \
+CUDA_VISIBLE_DEVICES=$2 python $this_dir/main_gdrn.py \
+    --config-file $CFG --num-gpus $NGPU --eval-only \
+    --opts MODEL.WEIGHTS=$CKPT INPUT.WITH_DEPTH=True TEST.USE_DEPTH_REFINE=True \
+    ${@:4}