omrajkumar / CarND-Traffic-Light-Detection-TF

TensorFlow Object Detection API Tutorial for Traffic Light Detection(TLD)

1. Set your environment

https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md

git clone https://github.com/tensorflow/models.git 

# From tensorflow/models
vi research/object_detection/g3doc/installation.md

export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim 

# For CPU
pip install tensorflow

sudo pip install pillow
sudo pip install lxml
sudo pip install jupyter
sudo pip install matplotlib

# From tensorflow/models/research/
protoc object_detection/protos/*.proto --python_out=.

python object_detection/builders/model_builder_test.py

2. Test a Model

• Save images to models/research/object_detection

• Open models/research/objection_detection/object_detection_tutorial.ipynb

• Edit TEST_IMAGE_PATHS

TEST_IMAGE_PATHS = [ os.path.join(PATH_TO_TEST_IMAGES_DIR, 'image{}.jpg'.format(i)) for i in range(1, N+1) ]

• Here is my result: A result of the jupyter notebook

3. Prepare train images

• Using_your_own_dataset.md

• Sample .yaml

- annotations:
  - {class: Green, x_width: 52.65248226950354, xmin: 130.4964539007092, y_height: 119.60283687943263,
    ymin: 289.36170212765956}
  - {class: Green, x_width: 50.156028368794296, xmin: 375.60283687943263, y_height: 121.87234042553195,
    ymin: 293.90070921985813}
  - {class: Green, x_width: 53.33333333333326, xmin: 623.6595744680851, y_height: 119.82978723404256,
    ymin: 297.7588652482269}
  class: image
  filename: sim_data_capture/left0003.jpg
...

• Make TFRecord files

# From models/research/object_detection
mkdir training
python tf_record_sim.py --output_path training/training_sim.record
python tf_record_real.py --output_path training/training_real.record

4. Training the Model

• Preparing models

cd training
mkdir data
cp ./ ./data
mkdir configs
cp ../samples/configs/* ./configs
mkdir models & cd models
# models/research/object_detection/g3doc/detection_model_zoo.md
# download
wget http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2017_11_17.tar.gz
wget http://download.tensorflow.org/models/object_detection/ssd_inception_v2_coco_2017_11_17.tar.gz
wget http://download.tensorflow.org/models/object_detection/faster_rcnn_inception_v2_coco_2017_11_08.tar.gz
wget http://download.tensorflow.org/models/object_detection/faster_rcnn_resnet50_coco_2017_11_08.tar.gz
# gunzip
gunzip ssd_mobilenet_v1_coco_2017_11_17.tar.gz
gunzip ssd_inception_v2_coco_2017_11_17.tar.gz
gunzip faster_rcnn_resnet50_coco_2017_11_08.tar.gz
gunzip faster_rcnn_inception_v2_coco_2017_11_08.tar.gz
# un-tar
tar -xvf ssd_mobilenet_v1_coco_2017_11_17.tar
tar -xvf faster_rcnn_inception_v2_coco_2017_11_08.tar
tar -xvf faster_rcnn_resnet50_coco_2017_11_08.tar
tar -xvf ssd_inception_v2_coco_2017_11_17.tar
# copy .pb files
cp faster_rcnn_inception_v2_coco_2017_11_08/*.pb models/faster_rcnn_inception_v2_coco.pb
cp faster_rcnn_resnet50_coco_2017_11_08/*.pb models/faster_rcnn_resnet50_coco.pb
cp ssd_inception_v2_coco_2017_11_17/*.pb models/ssd_inception_v2_coco.pb
cp ssd_mobilenet_v1_coco_2017_11_17/*.pb models/ssd_mobilenet_v1_coco.pb
# copy .ckpt files
cp faster_rcnn_inception_v2_coco_2017_11_08/model.ckpt.data-00000-of-00001 models/faster_rcnn_inception_v2_coco.ckpt.data-00000-of-00001
cp faster_rcnn_resnet50_coco_2017_11_08/model.ckpt.data-00000-of-00001 models/faster_rcnn_resnet50_coco.ckpt.data-00000-of-00001
cp ssd_inception_v2_coco_2017_11_17/model.ckpt.data-00000-of-00001 models/ssd_inception_v2_coco.ckpt.data-00000-of-00001
cp ssd_mobilenet_v1_coco_2017_11_17/model.ckpt.data-00000-of-00001 models/ssd_mobilenet_v1_coco.ckpt.data-00000-of-00001

cp faster_rcnn_inception_v2_coco_2017_11_08/model.ckpt.index models/faster_rcnn_inception_v2_coco.ckpt.index
cp faster_rcnn_resnet50_coco_2017_11_08/model.ckpt.index models/faster_rcnn_resnet50_coco.ckpt.index
cp ssd_inception_v2_coco_2017_11_17/model.ckpt.index models/ssd_inception_v2_coco.ckpt.index
cp ssd_mobilenet_v1_coco_2017_11_17/model.ckpt.index models/ssd_mobilenet_v1_coco.ckpt.index

cp faster_rcnn_inception_v2_coco_2017_11_08/model.ckpt.meta models/faster_rcnn_inception_v2_coco.ckpt.meta
cp faster_rcnn_resnet50_coco_2017_11_08/model.ckpt.meta models/faster_rcnn_resnet50_coco.ckpt.meta
cp ssd_inception_v2_coco_2017_11_17/model.ckpt.meta models/ssd_inception_v2_coco.ckpt.meta
cp ssd_mobilenet_v1_coco_2017_11_17/model.ckpt.meta models/ssd_mobilenet_v1_coco.ckpt.meta
# rm model dirs
rm -rf faster_rcnn_inception_v2_coco_2017_11_08
rm -rf faster_rcnn_resnet50_coco_2017_11_08
rm -rf ssd_inception_v2_coco_2017_11_17
rm -rf ssd_mobilenet_v1_coco_2017_11_17

Model name	Speed (ms)	COCO mAP[^1]	Outputs
ssd_mobilenet_v1_coco	30	21	Boxes
ssd_inception_v2_coco	42	24	Boxes
faster_rcnn_inception_v2_coco	58	28	Boxes
faster_rcnn_resnet50_coco	89	30	Boxes
faster_rcnn_resnet50_lowproposals_coco	64		Boxes
rfcn_resnet101_coco	92	30	Boxes
faster_rcnn_resnet101_coco	106	32	Boxes
faster_rcnn_resnet101_lowproposals_coco	82		Boxes
faster_rcnn_inception_resnet_v2_atrous_coco	620	37	Boxes
faster_rcnn_inception_resnet_v2_atrous_lowproposals_coco	241		Boxes
faster_rcnn_nas	1833	43	Boxes
faster_rcnn_nas_lowproposals_coco	540		Boxes

• Mapping Data: models/research/objection_detection/traffic_light_map.pbtxt

  item {
    id: 0
    name: 'Red'
  }

  item {
    id: 1
    name: 'Yellow'
  }

  item {
    id: 2
    name: 'Green'
  }

  item {
    id: 4
    name: 'Undefined'
  }

• Modify configs: models/research/objection_detection/training/configs/ssd_mobilenet_v1_coco.config

model {
  ssd {
    num_classes: 4
    box_coder {
      faster_rcnn_box_coder {
        y_scale: 10.0
        x_scale: 10.0
        height_scale: 5.0
        width_scale: 5.0
      }
    }
    matcher {
      argmax_matcher {
        matched_threshold: 0.5
        unmatched_threshold: 0.5
        ignore_thresholds: false
        negatives_lower_than_unmatched: true
        force_match_for_each_row: true
      }
    }
    similarity_calculator {
      iou_similarity {
      }
    }
    anchor_generator {
      ssd_anchor_generator {
        num_layers: 6
        min_scale: 0.2
        max_scale: 0.95
        aspect_ratios: 1.0
        aspect_ratios: 2.0
        aspect_ratios: 0.5
        aspect_ratios: 3.0
        aspect_ratios: 0.3333
      }
    }
    image_resizer {
      fixed_shape_resizer {
        height: 300
        width: 300
      }
    }
    box_predictor {
      convolutional_box_predictor {
        min_depth: 0
        max_depth: 0
        num_layers_before_predictor: 0
        use_dropout: false
        dropout_keep_probability: 0.8
        kernel_size: 1
        box_code_size: 4
        apply_sigmoid_to_scores: false
        conv_hyperparams {
          activation: RELU_6,
          regularizer {
            l2_regularizer {
              weight: 0.00004
            }
          }
          initializer {
            truncated_normal_initializer {
              stddev: 0.03
              mean: 0.0
            }
          }
          batch_norm {
            train: true,
            scale: true,
            center: true,
            decay: 0.9997,
            epsilon: 0.001,
          }
        }
      }
    }
    feature_extractor {
      type: 'ssd_mobilenet_v1'
      min_depth: 16
      depth_multiplier: 1.0
      conv_hyperparams {
        activation: RELU_6,
        regularizer {
          l2_regularizer {
            weight: 0.00004
          }
        }
        initializer {
          truncated_normal_initializer {
            stddev: 0.03
            mean: 0.0
          }
        }
        batch_norm {
          train: true,
          scale: true,
          center: true,
          decay: 0.9997,
          epsilon: 0.001,
        }
      }
    }
    loss {
      classification_loss {
        weighted_sigmoid {
          anchorwise_output: true
        }
      }
      localization_loss {
        weighted_smooth_l1 {
          anchorwise_output: true
        }
      }
      hard_example_miner {
        num_hard_examples: 3000
        iou_threshold: 0.99
        loss_type: CLASSIFICATION
        max_negatives_per_positive: 3
        min_negatives_per_image: 0
      }
      classification_weight: 1.0
      localization_weight: 1.0
    }
    normalize_loss_by_num_matches: true
    post_processing {
      batch_non_max_suppression {
        score_threshold: 1e-8
        iou_threshold: 0.6
        max_detections_per_class: 100
        max_total_detections: 100
      }
      score_converter: SIGMOID
    }
  }
}

• Train & Inference - ssd_mobilenet_v1

# Train
python train.py --logtostderr --train_dir=./training/models/ --pipeline_config_path=./training/configs/ssd_mobilenet_v1_coco.config
# Inference
python export_inference_graph.py --input_type image_tensor --pipeline_config_path ./training/configs/ssd_mobilenet_v1_coco.config --trained_checkpoint_prefix ./training/models/model.ckpt-400 --output_directory ./training/fine_tuned_model

5. Applying the Model

• Sample Image

References

Step by Step TensorFlow Object Detection API Tutorial

TensorFlow Object Detection API