🔎 Approach Summary

• Framework: MMDetection
• Backbone: Swin Transformer (Large) with Co-DETR head
• Training Time Constraint: 9 hours
• Key Insight: A smaller, high-signal subset of the data + a large backbone outperformed training on the full dataset with a small model

📎 Inference Notebook: Kaggle Link – feel free to upvote if you find it useful! 🔼

📊 Smart Data Subsampling

To stay within time limits while using Swin-L, I applied a signal-based filtering strategy:

1. Created a fixed validation set
2. Established baseline model trained on small seed
3. Iteratively added training samples in batches
4. Retained only those batches that improved validation score

This led to a compact and high-signal training set.

🚀 Model Pipeline

• Backbone: Swin-Large pretrained on ImageNet-22k
• Detector: Co-DETR with 5-scale deformable attention
• Training Augmentations: Color jitter Horizontal, vertical, diagonal flips Random resize to 512/640 px – Multi-scale training
• Training Epochs: 10
• Optimizer: AdamW (lr=1e-4, weight decay=1e-4)

💡 Postprocessing & Ensembling

• Checkpoints: Used last 3 epochs (8, 9, 10) for ensembling
• TTA: Horizontal flip, Vertical and Transpose
• Soft-NMS: Class-wise with IoU threshold 0.5
• Ensembling: Bayesian-style box fusion across TTA + checkpoints

🌐 Environment

• GPU: Tesla T4 (Kaggle)
• PyTorch: 2.5.1
• MMDetection: 3.3.0

📅 Reproducibility

• All code, config, and checkpoints are available in the shared Kaggle notebook
• Run all to get the submission file

Massive thanks to the Zindi team, Makerere AI Lab, and Amini.ai for this powerful real-world challenge! 🚀