Rank 1 @ Kaggle UltraMNIST

Independent Project — Kaggle UltraMNIST Competition (Date)

Overview
Secured 1st place on the Kaggle UltraMNIST challenge with 99.109% test accuracy. Combined a small-object detector (YOLOv5s) and a high-accuracy classifier (EfficientNetV2-B1) on a custom synthetic dataset to handle 2560×2560 images containing up to five tiny objects.

Key Components

• Synthetic Data Generation
  • Created checkerboards with boxes, circles, triangles. Used solvers to determine placement of digits so that they dont overlap.
  • Applied perspective, rotation, zoom, and uniform digit placement (all 70 000 MNIST digits + empty images).
• Detection (YOLOv5s)
  • Input: 2560×2560; augmentation: light rotate/scale/translate/HSV/mosaic.
  • Config: conf_threshold=0.1, iou_threshold=0.1, class-agnostic NMS, max_det=5.
  • Recall > 95%.
• Classification (EfficientNetV2-B1)
  • Input: 128×128 crops (aspect-ratio normalized).
  • Pretrained on ImageNet; light aug (shift/scale, perspective, invert, blur, HSV).
  • Training: 200 epochs, LR schedule [3e-4 → 1e-5], TTA (invert images).
  • Accuracy > 98%.
• Post-Processing
  • Removed nested detections and low-confidence (< 0.45) preds.
  • Final ensemble accuracy: 99.109%.

Resources

• Code & Models: github.com/4g/umnist
• Competition: UltraMNIST on Kaggle
• Writeup : Kaggle writeup