# [From 447 to 135th place Solution] EfficientDet Ensemble

Congratulations to all the winners and to all the players that learnt new things!
Since the start of the competition we have used FasterRCNN then Shifted to EfficientDet then Shifted to YOLOv5. Then used a combination of both EfficientDet+YOLOv5. When we knew that YOLOv5 is prohibited we turned back to EfficientDet and kept optimizing it.
Training
Data Cleaning:
Deleted tiny bounding boxes (width or height under a certain threshold)
Deleted hunge bounding boxes (width or height over a certain threshold)
Data Augmentation:
Online Augmentations
Flips & Rotations (90 multiples)
Crop & Resize
Hue Saturation Contrast Brightness
Moving Blur
Noise Injection (Gaussian Noise)
Cutout (Made different versions of that)
Cutmix
Training Process:
Used EfficientDet5 (& EfficientDet7)
Used both 512x512  and 1024x1024 images
Startified 5 Folds Splitting
By Source Splitting (for some Fine Tuning)
Manual tweaks when training the models (& Early Stopping)
Change Augmentation distributions every certain number of epochs
Inference
A picture is worth a thousand words:
![](assets/ec66b6af50b0a65c4f569cbc.png)
Here you find my inference
[kernel](https://www.kaggle.com/ffares/ensemble-models-tta-efficientdet-512-1024)
, please upvote the notebook if you like it, and most imprtantly feel free to give your feedback so I can learn from my mistakes.
I want to thank the hosts of this competition, kaggle and all the participants who submitted compliant solutions. Special thanks goes for
[@shonenkov](https://www.kaggle.com/shonenkov)
[@nvnnghia](https://www.kaggle.com/nvnnghia)
[@pestipeti](https://www.kaggle.com/pestipeti)
and many others who kept sharing during the competition.
We will update with further details.