86Cao/google-landmark-v2-chinese-filtered

Google Landmark V2 Chinese Filtered Dataset

This dataset contains landmark images and metadata for training landmark retrieval models, with Chinese translations of landmark names to facilitate Chinese multimodal retrieval tasks.

Dataset Source

This dataset is based on the Google Landmarks V2 dataset from Kaggle. The original data has been filtered and processed to create a high-quality training dataset for landmark retrieval.

Key Features

1. Filtered from Kaggle Dataset: The dataset is filtered from the cleaned subsets of Google Landmarks V2 available on Kaggle, ensuring data quality. train_filter.json and valid_filter.json contain the filtered training and validation sets respectively.

2. Chinese Translations for VLM-based Multimodal Retrieval: All landmark names have been translated to Chinese, making this dataset ideal for training Vision-Language Models (VLM) based multimodal retrieval systems. By providing Chinese text alongside images, VLMs can better leverage their rich pretrained knowledge to achieve significant performance improvements in landmark retrieval tasks.

3. Category-based Organization: Landmarks are organized by categories, enabling category-based hard negative sampling for improved training.

Dataset Structure

Metadata Files

• train_filter.json: Filtered training set metadata (36,493 landmarks, 615,419 query images)

• valid_filter.json: Filtered validation set metadata (36,493 landmarks, 151,051 query images)

• landmark_id_to_category.json: Complete mapping from landmark_id to category (includes all categories, 1,114 categories, 36,493 landmarks). This file contains all landmarks with valid category information, including some categories that may not be suitable for landmark recognition (e.g., administrative entities, abstract concepts, generic geographic terms). Use this file if you want to include all available category information.

• landmark_id_to_category_filtered.json: Filtered mapping (valid landmark categories only, 283 categories, 27,516 landmarks). This file only includes landmarks with categories that are suitable for landmark recognition, filtered using the same criteria as in tools/clean_poi_categories.py. Categories like organizations, administrative entities, generic geographic terms, etc. are excluded. Recommended for training as it ensures only high-quality landmark categories are used.

Image Folders

• train_filtered/: Training images organized by landmark_id
• valid_filtered/: Validation images organized by landmark_id

Dataset Statistics

Original Dataset (Before Category Filtering)

• Train: 36,493 landmarks, 615,419 query images
• Valid: 36,493 landmarks, 151,051 query images
• Total: 36,493 landmarks, 766,470 query images

Filtered Dataset (Valid Categories Only)

After filtering using landmark_id_to_category_filtered.json:

• Train: 27,516 landmarks, 463,873 query images (75.38% retention)
• Valid: 27,516 landmarks, 113,682 query images (75.26% retention)
• Total: 27,516 landmarks, 577,555 query images (75.35% retention)

Note: Approximately 25% of the data (8,977 landmarks) were filtered out as they belong to categories not suitable for landmark recognition (e.g., administrative entities, abstract concepts, generic geographic terms).

Data Format

JSON Metadata Format

Each landmark entry in train_filter.json and valid_filter.json has the following structure:

{
  "landmark_id": "116936",
  "landmark_name": "塔彭卡湖",
  "reference_image": "train_filtered/116936/014ccab2d16604e1.jpg",
  "query_images": [
    "train_filtered/116936/1f0102770a37d22f.jpg",
    "train_filtered/116936/21d82cf8c60fc770.jpg",
    ...
  ],
  "supercategory": "lake",
  "hierarchical_label": "lake"
}

Category Mapping Format

landmark_id_to_category_filtered.json maps landmark_id to category:

{
  "church": ["100005", "100017", ...],
  "castle / fort": ["100049", ...],
  ...
}

Categories are sorted by the number of landmarks (from most to least).

Usage

Loading the Dataset

import json

# Load training metadata
with open("train_filter.json", "r") as f:
    train_data = json.load(f)

# Load validation metadata
with open("valid_filter.json", "r") as f:
    valid_data = json.load(f)

# Load complete category mapping (all categories)
with open("landmark_id_to_category.json", "r") as f:
    category_mapping_all = json.load(f)

# Load filtered category mapping (valid categories only)
with open("landmark_id_to_category_filtered.json", "r") as f:
    category_mapping_filtered = json.load(f)

Category-based Hard Negative Sampling

This dataset supports category-based hard negative sampling, a training strategy that improves model performance by using more challenging negative examples:

What are Hard Negatives?

• Hard Negatives: Landmarks from the same category but different landmark_id

  • Example: If your query is from the "church" category (landmark_id: "100005"), other "church" landmarks (e.g., "100017", "100049") are treated as hard negatives
  • These are more challenging than random negatives because they share similar visual characteristics (e.g., architectural style, structure type)

• Regular Negatives: Landmarks from different categories

  • Example: A "church" query vs "castle" or "museum" landmarks are regular negatives
  • These are easier to distinguish as they have different visual characteristics

Training Strategy

When training landmark retrieval models, you can enable category-based hard negative sampling The loss function combines both hard negative loss and regular negative loss with a configurable weight:

loss = hard_negative_weight * loss_hard + (1 - hard_negative_weight) * loss_standard

This allows the model to:

• Learn fine-grained distinctions within categories (via hard negatives)
• Still benefit from diverse negative examples across categories (via regular negatives)

**Parameters:**
- `--category_mapping_path`: Path to the category mapping file (use `landmark_id_to_category_filtered.json` for filtered categories)
- `--hard_negative_weight`: Weight for hard negative loss (0.0-1.0)
  - `0.0`: Only use regular negatives (standard in-batch negative sampling)
  - `0.5`: Equal weight for hard and regular negatives
  - `0.7-0.8`: Recommended range, emphasizing hard negatives
  - `1.0`: Only use hard negatives (same-category negatives only)

#### Data Filtering

The category mapping can also be used to filter out landmarks without valid categories during training, ensuring only high-quality landmark data is used. Landmarks not present in `landmark_id_to_category_filtered.json` will be automatically skipped during data loading.

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{google_landmark_v2_chinese_filtered,
  title={Google Landmark V2 Chinese Filtered Dataset},
  author={Your Name},
  year={2025},
  url={https://huggingface.co/datasets/86Cao/google-landmark-v2-chinese-filtered}
}

License

Creative Commons Attribution 4.0 International (CC-BY 4.0)

This dataset is released under CC-BY 4.0, the same license as the original Google Landmarks V2 dataset. You are free to:

• Share: Copy and redistribute the material in any medium or format
• Adapt: Remix, transform, and build upon the material for any purpose, even commercially

Under the following terms:

• Attribution: You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use.

This license is compatible with the original Google Landmarks V2 dataset license, ensuring consistency with the source data.

For more information, see: https://creativecommons.org/licenses/by/4.0/