Naver AI Tech 7th

Overview

During the Naver AI Tech 7th (Computer Vision Track), I conducted four major projects ranging from image classification to segmentation. Through this intensive curriculum, I developed expertise in Data-Centric AI, Model Optimization, and MLOps tools (WandB, Poetry).

Below is a summary of the key projects and my specific contributions.

1. Sketch Image Classification (ImageNet-Sketch)

👉 Code

SketchDataSet_Example1

SketchDataSet_Example2

Objective Build a robust baseline image classification model on the ImageNet-Sketch dataset, focusing on improving generalization performance through data-centric strategies.

🎯 My Role:

• Led dataset analysis and data-centric optimization, identifying performance bottlenecks caused by dataset structure rather than model capacity
• Designed and validated data cleaning policies for duplicate and ambiguous samples
• Implemented and evaluated augmentation and preprocessing strategies tailored to sketch-style images
• Managed experiments and hyperparameter searches using Weights & Biases

Approach & Key Solutions

🛠️ Methods:

• Exploratory Data Analysis (EDA)
  • Analyzed image count per class and class imbalance
  • Identified near-duplicate sketches caused by flips and rotations
  • Examined object placement and background patterns to guide preprocessing decisions
• Data Cleaning
  • Removed near-duplicate samples dominating certain classes
  • Pruned ambiguous samples between visually overlapping classes (e.g., baseball icon vs. baseball player)
  • Verified that simple removal outperformed synthetic data generation approaches
• Targeted Augmentation (Albumentations)
  • Avoided Flip/Rotate-heavy augmentations already present in the dataset
  • Applied controlled ShiftScaleRotate (wrap mode) to improve spatial robustness
  • Used GaussianBlur to reduce overfitting to repeated line patterns
  • Empirically confirmed that aggressive augmentations (e.g., CoarseDropout, GridShuffle) degraded performance
• Grayscale-aware Preprocessing
  • Identified that ImageNet-Sketch images are inherently grayscale
  • Compared:
    1. Replicating grayscale to 3-channel input
    2. Modifying pretrained Conv2D weights for single-channel input
  • Computed dataset-specific mean and standard deviation for grayscale normalization
• Experiment Management
  • Tracked experiments and metrics using WandB
  • Performed hyperparameter optimization via WandB Sweeps
  • Compared CNN-based and ViT-based models from an inductive bias perspective

📊 Result:

• Achieved up to ~0.92 classification accuracy, outperforming naive augmentation baselines
• Demonstrated that dataset de-duplication and ambiguity removal provided larger gains than architectural changes
• Confirmed that indiscriminate augmentation degrades performance on already-distorted sketch data

Key Insight For sketch-based image classification, data quality and class clarity matter more than model complexity. Removing misleading samples consistently improved generalization, while indiscriminate augmentation led to performance collapse.

2. Trash Object Detection

👉 Code 👉 Report

Objective: Detect and classify trash in high-resolution (1024x1024) images into 10 categories.

• 🔍 EDA: ① Box ② Class ③ Color

• 🎯 My Role:
  • YOLO Implementation: Experimented with YOLOv11 (s, l, xl) models to verify the performance of 1-stage detectors compared to 2-stage models.
  • Ensemble Strategy: Contributed to the final ensemble by combining YOLO predictions with the team’s Cascade R-CNN and DINO models.
• 🛠️ Methods:
  • Augmentation: Applied Mosaic, MixUp, and RandomCrop to handle object scale variations.
  • Ensemble: Used WBF (Weighted Boxes Fusion) and Soft-NMS to merge bounding boxes effectively.
• 📊 Result: Achieved Public Score 0.6714 / Private Score 0.6558.
  You can find the detailed experimental results below.

• Summary

  Conducted large-scale object detection experiments using YOLOv11 and multi-model ensembles (Cascade R-CNN, DINO), achieving Public 0.6714 / Private 0.6558 through optimized augmentation and bounding-box fusion strategies.

3. Data-Centric: Multilingual Receipt OCR

👉 Code 👉 Report

OCR_Example1

OCR Example2

Objective: Improve OCR performance for multilingual receipts (Chinese, Japanese, Thai, Vietnamese) purely through Data Preprocessing (Model architecture fixed).

• Metrics: DetEval (F1 Score)
• 🎯 My Role:
  • Data Labeling & Cleaning: Manually corrected orientation issues and removed noise from the Ground Truth labels to improve data quality.
  • Visualization: Developed a Streamlit dashboard to visualize data distribution and augmentation effects for the team.
• 🛠️ Methods:
  • Dataset Expansion: Integrated external datasets (ICDAR, CORD).
  • Augmentation: Applied Perspective Transform and various noise injections (Gaussian, Salt & Pepper).

• 📊 Result: Achieved F1 Score 0.8100 (Public) / 0.8078 (Private).

• Performance Comparison (Before vs. After Ensemble)

• Summary

After applying an optimized ensemble strategy, the F1 score improved from a maximum of 0.7366 (single model) to 0.8100 (Public) and 0.8078 (Private), demonstrating a balanced improvement in both precision and recall.

4. Bone Segmentation

👉 Code 👉 Report

This Streamlit app was developed to provide detailed, step-by-step visualization of hand bone segmentation examples.

Objective: Develop a segmentation model to precisely identify bone areas in medical X-ray images.

• Metrics: Dice Score
• 🎯 My Role & Technical Approach:
  • Library Comparison & Data Workflow: Conducted analysis of folder structures and performed detailed data EDA, then refactored the pipeline for efficient data preprocessing and augmentation using various libraries.
  • Encoder Experiments: Analyzed performance differences between Transformer-based encoders and CNN-based encoders.
  • Troubleshooting: Resolved weight initialization errors (encoder_weights: None) and decoder compatibility issues.
• 🛠️ Methods:
  • Model Architecture: Found that U-Net provided more stable performance on medical data compared to DeepLabV3+ or U-Net++.
  • Hyperparameter Tuning: Optimized Image Size and Epochs using WandB Sweeps.
• 📊 Result:

Finally, the best result was obtained by combining high-resolution inputs with attention-enhanced decoding and a soft ensemble strategy, achieving a Dice score of 0.9751 on the test set.

📸 Commemorative Photos