Congrats to the winners and all participants and Thank you to Zindi and Digital Africa for setting up this challenge, it was a great opportunity to experiment with Grad-CAM, TestTimeAugmentation and various computer vision models and methodologies for the first time, I’ve learned so much, thank you.

Remarks:

Smaller image sizes and larger batch sizes I found to be the simple tricks to obtain better validation scores.

Use Test-Time-Augmentaion! Without Test-Time-Augmentaion (TTA) this approach would have scored a 1.622, which I think (due to my cropping augments) mainly indicates the sensitivity of the model to trees on the boarder of the image. On my validation set, TTA improved the score by 0.0167, which seems small, but on the public testset this translated to an improvement of 0.024, from 2.132 to 2.107.

While this single model setup scored an RMSE of 1.595 on the private testset, a similar process within stage 2 can be applied to the other non selected train/val splits from stage 1 to average more predictions and obtain better results (1.556). I have experimented with this but it was a tough choice to make as my final modelling pipeline because of resource contraints and because of the challenges within the dataset that created seemingly large deviations across the folds training and validation scores. This approach was not attractive. I rather choose to focus on one model and use the available resources to get more experiments in.

My solutions:

Stage 1: https://www.kaggle.com/code/danielbruintjies/stage-1-of-tree-counting-image-regression

Stage 2: https://www.kaggle.com/code/danielbruintjies/stage-2-of-tree-counting-image-regression

Please upvote them if you find it useful, I am also open to your critique and ways to improve, looking forward to seeing others approaches as well!