Boxy Keypoints Detection

This project used Boxy dataset from Bosch and Centernet from the Tensorflow Object Detection API to detect vehcles 3D box.

0. Install Environment

Step 1. Install Bosch Boxy Dataset

Download Dataset From : https://boxy-dataset.com/boxy/

Step 2. Install Object Detection API

Install Tensorflow Object Detection API : https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2.md

Step 3. Clone and cover

Clone this project and cover and overwrite the direction to object_detection

Step 4. Update protos

cd models/research
# Update protos.
protoc object_detection/protos/*.proto --python_out=.

1. Train Boxy

Step 1. Create Training Dataset

cd object_detection/dataset_tools/
python create_coco_tf_record.py  \
      --train_zip_dir="${TRAIN_ZIP_DIR}" \
      --train_annotations_file="${TRAIN_ANNOTATIONS_PATH/boxy_labels_train.json}" \
      --output_dir="${OUTPUT_DIR}"

Step 2. Modify TF record file in Configure file

Can pick a configure file for training

object_detection/configs/tf2/centernet_mobilenet_v2_fpn_512x512_kpts_boxy.config
object_detection/configs/tf2/centernet_resnet50_v2_512x512_kpts_boxy.config

Edited the path of the tfrecord path into the configure file you selected.

tf_record_input_reader {
    input_path: MODIFY_HERE
  }

Step 3. Train Model

Train the model by follewing command:

python model_main_tf2.py \
  --model_dir MODEL_DIR \
  --pipeline_config_path PIPELINE_CONFIG_PATH

2. Export SavedModel

Export to savemodel pb file from checkpoint:

python exporter_main_v2.py \
    --pipeline_config_path PIPELINE_CONFIG_PATH \
    --trained_checkpoint_dir CHECKPOINT_DIR \
    --output_directory OUTPUT_DIR

3. Inference Model

Inference model on image

python object_detection/detection_kpt_by_image.py \
    --model_path PATH_OF_MODEL \
    --image_path PATH_OF_VIDEO

Inference model on video

python object_detection/detection_kpt_by_video.py \
    --model_path PATH_OF_MODEL \
    --video_path PATH_OF_VIDEO

hhsinhan/boxy_tf2_keypoint_detection