Car plates OCR Competition Solution within MADE Computer Vision Course

Predict car plate numbers from the pictures.

https://www.kaggle.com/c/car-plates-ocr-made

General Approach

1. Detect the car plates

• MASKRCNN model with ResNet50 backbone for bounding box detection and mask segmentation predictions
• Adam Optimizer with Reduce On Plateau scheduler

2. Car plates OCR

• Generate a bounding box for each car plate and cut to form a rectangle
• CRNN model with ResNet18 backbone
• AdamW optimizer with Reduce On Plateau scheduler

Config

• TRAIN_SIZE = 0.9
• BATCH_SIZE = 4
• BATCH_SIZE_OCR = 128
• LR = 1e-3
• PATIENCE = 3
• FACTOR = 0.9
• THRESHOLD_SCORE = 0.93
• TRESHOLD_MASK = 0.05
• 1 epoch for detection and 3 epochs for OCR

Results

72nd Public and 67th Private place at the leaderboard!