/few-shot-detection

Few-shot traffic sign detection and classification

Primary LanguageJupyter Notebook

Few-shot traffic sign detection and classification

Roadmap:

  • Implement baseline non-few-shot traffic sign classification.
    • Parse classification part of Russian Traffic Sign Dataset
    • Implement basic training and testing routine
  • Extend classifier for heatmap generation
    • Baseline implementation
    • Improving with own ideas
    • Improving with ideas from paper
TBD by 22.03.2018
  • Sample own classification dataset from full RTSD dataset
    • Include resampling for most/least frequent classes
    • Include data augmentation by resizing signs and adding background
  • Create visualisations
    • Heatmap visualizations
    • Best/worst performing classes, confusion matrix
    • Heatmaps for full scenes
TBD by 1.04.2018
  • Apply classifier base for detection
    • Baseline implementation using fixed bounding boxes around heatmap maxima
    • Baseline implementation using already made detection networks