chilly-magician's picture
Update README.md
7c0d9a1
metadata
license: apache-2.0
dataset_info:
  features:
    - name: image
      dtype: image
    - name: category
      dtype: string
    - name: img_id
      dtype: string
  splits:
    - name: train
      num_bytes: 687610836.528
      num_examples: 26872
    - name: test
      num_bytes: 178694171.287
      num_examples: 6719
  download_size: 843239857
  dataset_size: 866305007.815
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
task_categories:
  - image-classification
tags:
  - landscapes
  - geo
  - remote photos
  - metric learning
pretty_name: Merged Remote Landscapes v1.0.0
size_categories:
  - 10K<n<100K

Dataset Card for Merged Remote Landscapes dataset

version

Dataset summary

This is a merged version of following datasets:

from datasets import load_dataset

dataset = load_dataset('EmbeddingStudio/merged_remote_landscapes_v1')

Categories

This is a union of categories from original datasets: agricultural, airplane, airport, baseball diamond, basketball court, beach, bridge, buildings, chaparral, church, circular farmland, cloud, commercial area, desert, forest, freeway, golf course, ground track field, harbor, industrial area, intersection, island, lake, meadow, mountain, overpass, palace, parking lot, railway, railway station, rectangular farmland, residential, river, roundabout, runway, sea ice, ship, snowberg, stadium, storage tanks, tennis court, terrace, thermal power station, wetland

Warning: Synonymous and ambiguous categories were combined (see "Merge method").

Motivation

EmbeddingStudio is the open-source framework, that allows you transform a joint "Embedding Model + Vector DB" into a full-cycle search engine: collect clickstream -> improve search experience-> adapt embedding model and repeat out of the box.

In the development of EmbeddingStudio the scientific approach is a backbone. On the early stage of the development we can't collect real clickstream data, so to do experiments and choose the best way to improve embedding model we had to use synthetic or emulated data. And the first step is to use the most transparent datasets and the easiest domain.

P.S. this dataset is tagged to be used for the image classification task, but in fact we use it for the metric learning task. And we do another step to emulate clickstream.

We provide this dataset on HuggingFace, so anyone can reproduce our results.

Check our repositories to get more details:

  • EmbeddingStudio Framework (coming soon at 22.12.2023)
  • Experiments (coming soon)

Merge method

For this type of dataset it's all simple:

  1. Remove duplicates.
  2. Resolve synonymous and ambiguous categories with using a simple map (CATEGORIES_MAPPING).
CATEGORIES_MAPPING = {
"dense residential": "residential",
"medium residential": "residential",
"mobile home park": "residential",
"sparse residential": "residential",
"storage tank": "storage tanks",
"storage tanks": "storage tanks",
}

All details and code base of merging algorithm will be provided in our experiments repository. If you have any suggestion or you find some mistakes, we will be happy to fix it, so our experimental data will have better quality.

Contact info