Taewan Kim, The University of Texas at Austin.
This is a code for detecting pedestrian with depth prediction using Deep Learning. TensorFlow Object Detection API is used for the developement.
If your current path is CURR, please download tensorflow/models repo (clone repo) as CURR/models, and follow the instructions for installing object detection API. Our repo should be placed as CURR/svdped to meet the path dependencies.
First, training, (validation), and test dataset should be prepared. Each set consists of image, lidar, and label per frame. Visit our carstop_ped repository to learn details about the structure of dataset and how to generate.
We used SSD Inception V2 (Batch Normalized Inception) pretrained on MS COCO which can be downloaded from SSD_Inception_v2_COCO (Tensorflow Object Detection Model Zoo). Please upzip the file at CURR/svdped/pretrained so that frozen graph can be found at CURR/svdped/pretrained/ssd_inception_v2_coco_2017_11_17/frozen_inference_graph.pb.
As recommended by TensorFlow, we also used TFRecords file to train and evaluate our model. Run create_cstopp_tfrecord.py code to transform prepared dataset to TFRecords. You can also use a script gen_cstopp_tfrecords.sh in the scripts subfolder.
Run train_ssd_inception_v2_org2.sh
Here we have several options. (1) just train, (2) use partial training (3) use partial training with initial max pooling (IMP)
-
For training a model without any options, refer to a script train_ssd_inception_v2_carstop.sh
-
For training a model with partial training (first train only lidar-related parts and box predictors), refer to a script partial_train_ssd_inception_v2_carstop.sh
-
For training a model with partial training and IMP, refer to a script imp_partial_train_ssd_inception_v2_carstop.sh
-
Without distance (no finetuning): infereval_ssd_inception_v2_plain.sh
-
Without distance: export_infereval_ssd_inception_v2_org2.sh
-
With distance (plain): export_infereval_ssd_inception_v2_carstop.sh
-
With distance + partial training: partial_export_infereval_ssd_inception_v2_carstop.sh
-
With distance + partial training + IMP: imp_partial_export_infereval_ssd_inception_v2_carstop.sh