1st place solution
#21
by
lqdisme
- opened
Congrats to all the winners, and thanks to organizers.
Summary
My solution is quite simple and totally based on YOLOv5, I don't have enough time to invest in another version of YOLO.
Link to full code pipeline: https://github.com/quangdungluong/hf-ship-detection
Link to sample training notebook: https://www.kaggle.com/lqdisme/hf-ship-detection
Model training
Data
I cut the original image into many patches with the size 400x400 and stride 300, then only trained YOLOv5 on those patches with ships (+ 10% random patches with no ships)
Validation strategy:
Split 5 folds
Models
YOLOv5s, YOLOv5m
Training details
Training and Inference with the image size 1024 x 1024, using default hyperparameters. (No TTA)
Ensemble
I ensemble many experiments using WBF