CaMiT / README.md
fredericlin's picture
Update README.md
27f7684 verified
metadata
license: cc-by-nc-sa-4.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
      - split: pretrain
        path: data/pretrain-*
dataset_info:
  features:
    - name: id
      dtype: string
    - name: time
      dtype: string
    - name: metadata
      struct:
        - name: id
          dtype: string
        - name: owner
          dtype: string
        - name: title
          dtype: string
        - name: license
          dtype: string
        - name: dateupload
          dtype: string
        - name: tags
          dtype: string
        - name: url_z
          dtype: string
        - name: height_z
          dtype: string
        - name: width_z
          dtype: string
        - name: date
          dtype: string
    - name: faces
      sequence:
        - name: bbox
          sequence: int32
        - name: det_score
          dtype: float32
    - name: boxes
      sequence:
        - name: bbox
          sequence: float32
        - name: yolo_score
          dtype: float32
        - name: class
          dtype: string
        - name: gpt_car_probability
          dtype: float32
        - name: gpt_model_probability
          dtype: float32
        - name: gpt_student_score
          dtype: float32
        - name: qwen_student_score
          dtype: float32
  splits:
    - name: train
      num_bytes: 294865622
      num_examples: 655681
    - name: test
      num_bytes: 38698484
      num_examples: 84830
    - name: pretrain
      num_bytes: 1207382492
      num_examples: 2709837
  download_size: 809497488
  dataset_size: 1540946598

CaMiT: Car Models in Time

CaMiT (Car Models in Time) is a large-scale, fine-grained, time-aware dataset of car images collected from Flickr. It is designed to support research on temporal adaptation in visual models, continual learning, and time-aware generative modeling.

Dataset Highlights

  • Labeled Subset:

    • 787,000 samples
    • 190 car models
    • 2007–2023
  • Unlabeled Pretraining Subset:

    • 5.1 million samples
    • 2005–2023
  • Metadata includes:

    • Image URLs (not the images themselves)
    • Upload time
    • Bounding boxes
    • Tags and licensing
    • Face detections
    • Car detection scores and class information

All images are linked (not redistributed) using Flickr metadata to respect copyright compliance, in line with LAION and DataComp datasets.

Tasks Supported

  • Time-aware fine-grained classification
  • Time-incremental continual learning
  • In-domain static and incremental pretraining
  • Time-aware image generation

Example Use Cases

  • Evaluating representation drift over time
  • Training classifiers that generalize across time periods
  • Studying model degradation and adaptation across years
  • Conditioning generation models on temporal context

Related Resource

A separate dataset containing precomputed image embeddings, categorized by car class and year, is available here:

👉 CaMiT Embeddings

License

CC BY-NC-SA 4.0 – for non-commercial research purposes only.