Model Files and Training Configuration

DetectorPlacas ships with a pre-trained SSD MobileNet v1 model and a pipeline.config file that fully describes how that model was trained. Every value documented on this page comes directly from the actual files in the repository — no defaults or assumptions have been added. Understanding these artifacts helps you evaluate the model’s characteristics and re-train or fine-tune it for your own dataset.

Model Artifacts

The following files make up the complete model bundle. All detection scripts load frozen_inference_graph.pb at runtime; the checkpoint files are needed only for fine-tuning.

File	Purpose
`inference_graph/frozen_inference_graph.pb`	Primary inference artifact. Loaded by all three detection scripts via `tf.gfile.GFile` and `tf.import_graph_def`.
`inference_graph/model.ckpt.data-00000-of-00001`	Checkpoint variable data — the actual weight values.
`inference_graph/model.ckpt.index`	Index file mapping variable names to their offsets in the data file.
`inference_graph/model.ckpt.meta`	MetaGraph definition describing the graph structure and collections.
`saved_model/saved_model.pb`	SavedModel format export for serving or TF 2 compatibility layers.
`checkpoint`	Plain-text pointer file: `model_checkpoint_path: "model.ckpt"` and `all_model_checkpoint_paths: "model.ckpt"`.
`labelmap.pbtxt`	Label map defining the single detection class (`license_plate`, id: 1).

Pipeline Configuration

The sections below break down every major block in pipeline.config. Expand each accordion to see the configuration values and a corresponding snippet.

Model Architecture

The model uses the SSD (Single Shot MultiBox Detector) framework with a MobileNet v1 feature extractor. The input is fixed at 300×300 pixels.

Parameter	Value
Architecture	SSD
`num_classes`	1
Image resizer	`fixed_shape_resizer` — 300×300
Feature extractor type	`ssd_mobilenet_v1`
`depth_multiplier`	1.0
`min_depth`	16
Activation	`RELU_6`
Batch norm decay	0.9997
Batch norm epsilon	0.001
L2 regularizer weight	4e-05
Initializer	`truncated_normal_initializer` (mean=0.0, stddev=0.03)

model {
  ssd {
    num_classes: 1
    image_resizer {
      fixed_shape_resizer {
        height: 300
        width: 300
      }
    }
    feature_extractor {
      type: "ssd_mobilenet_v1"
      depth_multiplier: 1.0
      min_depth: 16
      conv_hyperparams {
        regularizer {
          l2_regularizer {
            weight: 4e-05
          }
        }
        initializer {
          truncated_normal_initializer {
            mean: 0.0
            stddev: 0.03
          }
        }
        activation: RELU_6
        batch_norm {
          decay: 0.9997
          center: true
          scale: true
          epsilon: 0.001
          train: true
        }
      }
    }
  }
}

Box Coder

The faster RCNN box coder scales the encoded bounding-box deltas so that the regression targets have a magnitude suitable for training.

Parameter	Value
Coder type	`faster_rcnn_box_coder`
`y_scale`	10.0
`x_scale`	10.0
`height_scale`	5.0
`width_scale`	5.0

box_coder {
  faster_rcnn_box_coder {
    y_scale: 10.0
    x_scale: 10.0
    height_scale: 5.0
    width_scale: 5.0
  }
}

Matcher

The argmax matcher assigns ground-truth boxes to anchors based on IoU overlap, using symmetric matched and unmatched thresholds.

Parameter	Value
Matcher type	`argmax_matcher`
`matched_threshold`	0.5
`unmatched_threshold`	0.5
`ignore_thresholds`	false
`negatives_lower_than_unmatched`	true
`force_match_for_each_row`	true

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
  }
}

Anchor Generator

The SSD anchor generator creates default boxes across 6 feature map layers, with scale ranging from 0.2 to 0.95 and five aspect ratios per location.

Parameter	Value
`num_layers`	6
`min_scale`	0.2
`max_scale`	0.95
Aspect ratios	1.0, 2.0, 0.5, 3.0, 0.3333

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
  }
}

Post-Processing

After the network produces raw scores and boxes, post-processing applies Non-Maximum Suppression and converts scores with a sigmoid function.

Parameter	Value
`score_threshold`	1e-08
`iou_threshold`	0.6
`max_detections_per_class`	100
`max_total_detections`	100
Score converter	`SIGMOID`

post_processing {
  batch_non_max_suppression {
    score_threshold: 1e-08
    iou_threshold: 0.6
    max_detections_per_class: 100
    max_total_detections: 100
  }
  score_converter: SIGMOID
}

Loss

The model uses separate loss functions for bounding-box regression and class prediction, with a hard example miner to focus training on difficult samples.

Parameter	Value
Localization loss	`weighted_smooth_l1`
Classification loss	`weighted_sigmoid`
`num_hard_examples`	3000
Hard example miner `iou_threshold`	0.5
Hard example miner `loss_type`	`CLASSIFICATION`
`max_negatives_per_positive`	3
`min_negatives_per_image`	0
`classification_weight`	1.0
`localization_weight`	1.0

loss {
  localization_loss {
    weighted_smooth_l1 {
    }
  }
  classification_loss {
    weighted_sigmoid {
    }
  }
  hard_example_miner {
    num_hard_examples: 3000
    iou_threshold: 0.5
    loss_type: CLASSIFICATION
    max_negatives_per_positive: 3
    min_negatives_per_image: 0
  }
  classification_weight: 1.0
  localization_weight: 1.0
}

Training Configuration

Training uses the Adam optimizer with a manual learning-rate step schedule. The model is fine-tuned from an existing checkpoint rather than trained from scratch.

Parameter	Value
`batch_size`	16
Data augmentation	`random_horizontal_flip`, `ssd_random_crop`
Optimizer	Adam
`num_steps`	200,000
`fine_tune_checkpoint`	`inference_graph/model.ckpt`
`from_detection_checkpoint`	true
`load_all_detection_checkpoint_vars`	true

Learning rate schedule:

Step range	Learning rate
0 → 4,500	0.0002
4,500 → 7,000	0.0001
7,000 → 10,000	8e-05
10,000+	4e-05

train_config {
  batch_size: 16
  data_augmentation_options {
    random_horizontal_flip {
    }
  }
  data_augmentation_options {
    ssd_random_crop {
    }
  }
  optimizer {
    adam_optimizer {
      learning_rate {
        manual_step_learning_rate {
          initial_learning_rate: 0.0002
          schedule {
            step: 4500
            learning_rate: 0.0001
          }
          schedule {
            step: 7000
            learning_rate: 8e-05
          }
          schedule {
            step: 10000
            learning_rate: 4e-05
          }
        }
      }
    }
    use_moving_average: false
  }
  fine_tune_checkpoint: "inference_graph/model.ckpt"
  from_detection_checkpoint: true
  num_steps: 200000
  load_all_detection_checkpoint_vars: true
}

Evaluation

Evaluation is run against a held-out set of 1,100 examples using COCO detection metrics.

Parameter	Value
`num_examples`	1100
`metrics_set`	`coco_detection_metrics`
`use_moving_averages`	false

eval_config {
  num_examples: 1100
  metrics_set: "coco_detection_metrics"
  use_moving_averages: false
}

The train_input_reader and eval_input_reader sections in pipeline.config contain Windows absolute paths hardcoded to the original developer’s machine (e.g. C:/Users/OKTOPUZ-SLID/Desktop/detector-placas/data/). Before using this config to retrain or evaluate the model, update both label_map_path and input_path values to point to the correct locations on your own system.

Get Started

Guides

Reference

Model Files and Training Configuration

Model Artifacts

Pipeline Configuration

Build docs developers (and LLMs) love

Get Started

Guides

Reference

Documentation Index

​Model Artifacts

​Pipeline Configuration

Build docs developers (and LLMs) love

Model Artifacts

Pipeline Configuration