Training and inference source code: https://github.com/quanvuhust/cgiar-eyes-on-the-ground-challenge

SPLIT DATA

I split train data to 5 fold using group kfold. Group by prefix in image name.

REGRESION WITH BCE LOSS

Follow kaggle competition petfinder-pawpularity-score, I train model classification convnextv2_large with BCE loss, label = extent/100. I train only one fold 0 because training time is limited in 8h. It is risk when using only one fold, ensemble of 5 fold can boost private score.

SOFT LABEL

I use soft label because some images have wrong label.

After train model with BCE loss, I gen soft label with ratio: 0.7*gt + 0.3*predict

Final submission is ensemble by 3 model of fold 0, PB 9.4 and Private 10.598.

+ convnextv2_large orignal label, fold 0 epoch 6, image size 384

+ convnextv2_large soft label, image size 384

+ convnextv2_large soft label, image size 448

EXPERIMENTS

- ensemble two fold, fold 0 + fold 1, soft label, multiscale

- I have experiment with caformer_b36_384_in21ft1k, dinov2 vit base, dinov2 vit large, swinv2_large_window12to24_192to384.ms_in22k_ft_in1k, convnextv2_large.fcmae_ft_in22k_in1k_384. It hard to achieve 8.x public without using dataleak. Ensemble more model boost score, however increase training time.

- I do some experiment to test leak data:

+ coat_lite_medium_384.in1k with metadata, using damage column and growth_stage column as feature , concat metadata feature and image feature.. Public: 9.4 and Private 9.54 with coat lite medium :))

+ Train model classify 2 class with swin large using Test.csv as validation (achieve accuracy 92.5 on Test.csv). DR as class 1, Non DR as class 0. Set all test image has threshold < 0.5, 0.6, 0.7, 0.8, 0.9 to extent =0, ensemble with regression result. Small improve from public 9.58 -> 9.55

+ Compare result when training with all data and result using data leak (6.x public): RMSE between two predict (non DR test image) is ~ 7. Some test image has ground truth extent = 0, however model predict extent ~ 30, 40, 50, 60. I have visualized those images, they have wrong label. Some people can re assign value of model predict on such image to zeros (predicted extent = 0), this can boost RMSE score :))