FPT data centric competition
Introduction
Deep Learning models have become exceedingly developed and popular in recent years. On the other hand, data processing techniques have not been equally developed compared to models
In this competition, participants are provided with a dataset. The goal is to use processing techniques on that dataset to ensure that model achieves the best performance after training.
Following Reinforcement Learning Competition 2021 success, DataComp is a brand new competition with a new approach for researchers. Besides that, DataComp was created to contribute to the prevention of Covid-19 pandemic, using face mask recognition model.
- Competition link: https://datacomp.io/gioi-thieu
Our performance
- Achieve top 20/400 teams (5% highest team) having the highest score validated on the private test dataset
- Our [email protected] score on private test: 0.545
- Team name: "nan"
- Leaderboard link: https://datacomp.io/bang-xep-hang-cuoi-cung
Methods
We tried many different data augmentation from the basic types such as rotation, shearing, ... to some quite advance techniques such as mosaic, random safe crop,... The library that we're using albumentation
Consequently, the combination of these below technqiues result to the final highest score in our case:
- Train dataset ->
934 images
after relabeled to make sure the correctness is more than 99% - Validation dataset ->
154 images
(design an as much as general set by ultilizing KNN technique which is explained below!) - toGray augmentation ->
100 images
- CutOut + HorizontalFlip (p=0.5) ->
400 images
- Filter only incorrect-mask label images + HorizontalFlip (p=0.7) ->
200 images
- Mosaic augmentation ->
451 images
(Note: after do the mosaic augmentation, it's crucial to check the set again to exclude all images having poor-quality bboxes at the edge of each image) - Rotation + Shear (prob 50/50) ->
600 images
- Rotation + Shear (prob 50/50) with no-mask & mask only ->
200 images
- Remaining images augmented normally ->
400 images
- Rotation + Shear (prob 50/50) with no-mask & mask only ->
- B.c model perform poorly with images having people appeared behide the door. Therefore, filter & augment specificailly those images in training dataset ->
100 images
--> TOTAL 2939 augmentation images
to submit (training + validation)
KNN ultilization
- Briefly instroduce about KNN
- The application of KNN in our solution
- Used to construct as general as possible validation dataset
- Categorize type of images in training set to faster filter images with specific feature, characteristic (Ex: Img having people behide doors, img having people wearing different types of masks)