Danbooru2018 Anime Character Recognition Dataset

Danbooru 2018 Anime Character Recognition Dataset Overview

Dataset Description

• Dataset Name: Danbooru 2018 Anime Character Recognition Dataset
• Dataset Source: Based on the Danbooru 2018 dataset.
• Dataset Content: Contains 1,000,000 head images and their corresponding 70,000 character labels.
• Dataset Purpose: Used for training and evaluating anime character recognition algorithms.

Data Processing Method

• Label Filtering: Keep only character category labels.
• Image Filtering: Retain images that contain only a single character label.
• Head Detection: Use a specific model to extract head bounding boxes.
• Image Deduplication: Remove images with multiple detected head bounding boxes.
• Final Data Volume: 0.97M images, 70k labels.

Data Distribution and Visualization

• Label‑Image Count Distribution: Visualized, showing only the top 100 labels.
• Top 20 Popular Labels: Include hatsune_miku, hakurei_reimu, etc.
• Distribution Characteristics: Long‑tail distribution, average 13.85 images per label.

Dataset Usage

• Core Data File: faces.tsv, containing filename, label ID, and head detection results.
• Label Text File: tagIds.tsv, providing text for each label ID.
• Face Image Download: Pre‑processed face image archive can be downloaded via rsync.

Citation Information

• Dataset Author: Yan Wang
• Release Date: July 2019
• Citation Format: Please refer to the README for the BibTeX format.

Baseline Model

• Model Description: ResNet18 combined with ArcFace loss, achieving 37.3% accuracy.
• Data Split: Training, validation, and test split files are provided.

Open Issues

• Test Set Validation: Test set requires manual verification.
• Face Alignment: Further optimization of face alignment is needed.