RT-DETR Russian car plate detection with classification by type

This model is a fine-tuned version of PekingU/rtdetr_r50vd_coco_o365 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 4.1673
• Map: 0.8829
• Map 50: 0.9858
• Map 75: 0.9736
• Map Car-plates-and-these-types: -1.0
• Map Large: 0.9689
• Map Medium: 0.9125
• Map N P: 0.857
• Map P P: 0.9087
• Map Small: 0.696
• Mar 1: 0.8686
• Mar 10: 0.9299
• Mar 100: 0.9357
• Mar 100 Car-plates-and-these-types: -1.0
• Mar 100 N P: 0.9169
• Mar 100 P P: 0.9545
• Mar Large: 0.9844
• Mar Medium: 0.958
• Mar Small: 0.8354

Model description

Модель детекции номерных знаков автомобилей РФ, в данный момент 2 класса n_p и p_p, обычные номера и полицейские

Intended uses & limitations

Пример использования:

from transformers import AutoModelForObjectDetection, AutoImageProcessor
import torch
import supervision as sv


DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = AutoModelForObjectDetection.from_pretrained('Garon16/rtdetr_r50vd_russia_plate_detector').to(DEVICE)
processor = AutoImageProcessor.from_pretrained('Garon16/rtdetr_r50vd_russia_plate_detector')

path = 'path/to/image'
image = Image.open(path)
inputs = processor(image, return_tensors="pt").to(DEVICE)
with torch.no_grad():
    outputs = model(**inputs)
w, h = image.size
results = processor.post_process_object_detection(
    outputs, target_sizes=[(h, w)], threshold=0.3)
detections = sv.Detections.from_transformers(results[0]).with_nms(0.3)
labels = [
    model.config.id2label[class_id]
    for class_id
    in detections.class_id
]

annotated_image = image.copy()
annotated_image = sv.BoundingBoxAnnotator().annotate(annotated_image, detections)
annotated_image = sv.LabelAnnotator().annotate(annotated_image, detections, labels=labels)
  
grid = sv.create_tiles(
  [annotated_image],
  grid_size=(1, 1),
  single_tile_size=(512, 512),
  tile_padding_color=sv.Color.WHITE,
  tile_margin_color=sv.Color.WHITE
)
sv.plot_image(grid, size=(10, 10))

Training and evaluation data

Обучал на своём датасете - https://universe.roboflow.com/testcarplate/russian-license-plates-classification-by-this-type

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 300
• num_epochs: 20

Training results

Framework versions

• Transformers 4.46.0.dev0
• Pytorch 2.5.0+cu124
• Tokenizers 0.20.1