Data Centric for Noisy Animal Image Classification

Applying various data engineering techniques into image classification task for KAIST DS801 term project

1. What we’ve done so far

   • Outlier detection:
     • Based on the distances between image embedding and its class' average
     • Based on the distance towards the cluster centroid from 10-means clustering of the embeddings
   • Handling noisy labels:
     • Combining sample selection and relabeling using pre-trained EfficientNet
     • Fitted accumulated losses on 2 GMM and remove images that have 95% confidence coming from the larger loss (noise) distribution
     • Perform relabeling only when confidence > 90%
     • Final selection using confidence from ResNet-34
   • Tackling data imbalances
     • Using focal loss when handling noisy labels (pre-training, accumulating loss for sample selection, and when training for relabeling)
     • Augmentation only on images with high confidence from ResNet-34 using AutoAugment (ImageNet) and Random Cutmix + Mixup

2. Links

   • Google colab working paper
   • The dataset used is a modified version of the noisy ANIMAL-10N dataset