slamanigg/tiny-presence-detection

Use Case

This dataset is useful for demonstrations of how to train tiny vision models. The size of the dataset was reduced from ~25GB to ~300MB/900MB (grayscale/color) for quick and easy installation. This dataset focuses of the binary task of detecting a person in the image (similar to Visual Wake Words) but still contains all labels (needs to be manually relabelled to use for binary task).

Dataset Description

This dataset is a derivative work based on the Microsoft COCO (Common Objects in Context) Dataset created by the COCO Consortium. It modifies the original dataset by [briefly describe your changes, applying the preprocessing described below.

Preprocessing

1. Official COCO 2017 dataset.
2. Removal of all images tagged as "person" that do not have bounding boxes provided.
3. Boundig-box-aware cropping of the images to 1:1 aspect ratio (prevents cutting out the target information).
4. Downsampling to 64x64.
5. Transformation to grayscale (pixel dtype = uint8).
6. Removal of all iamges tagged as "person" that have small bounding boxes (area < 32x32 px).
7. Compression with GZIP.

NOTE: The "easy" version removes all images tagged as "person" that have bounding boxes with area < int(0.85*TARGET_SIZE)**2. Also the images are downsampled to 32x32 for fast demonstrations.

Size

Total images in raw COCO dataset (train+test+val): 163,957
Images discarded: 87,856
Images remaining: 76,101

How To Use

import os
import tensorflow as tf
from huggingface_hub import snapshot_download

# Define the repository details
REPO_ID = "slamanigg/tiny-presence-detection"

# Download into a local cache directory
repo_dir = snapshot_download(repo_id=REPO_ID, repo_type="dataset")
print(f"Repository files successfully downloaded to: {repo_dir}\n")

# Pick a version and load
CHOSEN_VERSION = "tiny_presence_64_CH_1"
version_path = os.path.join(repo_dir, CHOSEN_VERSION)
loaded_ds = tf.data.Dataset.load(version_path,bcompression='GZIP')
print(loaded_ds)

Limitations

This dataset was exported in tensorflow (2.20.0) and can therefor only be loaded with tensorflow.

Citation

If you use this dataset, please cite the following original COCO dataset papers:

@inproceedings{Lin2014COCO,
  author    = {Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll{\'a}r, Piotr and Zitnick, C. Lawrence},
  title     = {Microsoft COCO: Common Objects in Context},
  booktitle = {Computer Vision -- ECCV 2014},
  year      = {2014},
  pages     = {740--755},
  publisher = {Springer}
}

@article{Chen2015COCOCaptions,
  title   = {Microsoft {COCO} Captions: Data Collection and Evaluation Server},
  author  = {Chen, Xinlei and Fang, Hao and Lin, Tsung-Yi and Vedantam, Ramakrishna and Gupta, Saurabh and Doll{\'a}r, Piotr and Zitnick, C. Lawrence},
  journal = {arXiv preprint arXiv:1504.00325},
  year    = {2015}
}