ecopus/sign_identification

Dataset Card for Dataset Name

This dataset is designed for binary image classification of stop signs. It contains two splits: 'original' with 35 images and 'augmented' with 385 images. The augmented split was created by applying various image transformations to the original images to increase the dataset size and diversity for model training.

Dataset Details

Dataset Description

This dataset contains images of signs with binary labels (class_0 == image has stop sign and class_1 == image doesn't have a stop sign). The 'original' split contains the raw images, while the 'augmented' split contains augmented versions of the images generated using various transformations like random resized crop, horizontal/vertical flips, rotation, color jitter, sharpness adjustment, and autocontrast.

• Curated by: Emily Copus
• Shared by: @ecopus (Hugging Face Hub)
• Language(s) (NLP): English
• License: apache-2.0

Dataset Sources [optional]

• Repository: https://huggingface.co/datasets/ecopus/sign_identification

Uses

Direct Use

This dataset can be used for training and evaluating image classification models for identifying the two classes of signs. The augmented split can be particularly useful for improving model generalization and robustness, especially with a limited number of original examples.

Out-of-Scope Use

This is a limited dataset, with a single, binary target assignment per image. Ultimately, this dataset is not suited to contribute to ML algorithms outside of the identification of specifically stop signs. The user should refrain from utilizing models trained with this data for sensitive stop sign identification applications (i.e., self-driving applications).

Dataset Structure

The dataset has two splits: 'original' and 'augmented'. Each split contains two features:

• image: A datasets.Image object representing the image.
• label: A datasets.ClassLabel object with two classes: 'class_0' and 'class_1'. Class_0 refers to all images in which a stop sign is not present - conversely, class_1 refers to all images in which a stop sign is present. The 'original' split contains 35 examples and the 'augmented' split contains 385 examples.

Dataset Creation

Curation Rationale

This dataset was curated as a basic learning tool for implementing ML tools with image datasets. The simplicity of this dataset allows for easy implementation into basic ML binary classification algorithms, pefect for a first time user.

Source Data

The images in this dataset were manually captured by the curator from non-disclosed locations.

Data Collection and Processing

Again, this data was collected and compiled manually. The criterion for an acceptable image was: (a) contains some road sign (b) the full sign can be seen in the image (c) any text on the sign can be distinguished by the naked eye Class_1 (stop sign is in the image) consists of about 50% of the original split.

Who are the source data producers?

This data was produced entirely by the curator.

Annotation process

Labels (1 or 0) were manually assigned to each image at their respective indices.

Who are the annotators?

This dataset was annnotated by the curator.

Personal and Sensitive Information

This dataset contains no personal or sensitive information.

Bias, Risks, and Limitations

Technical Limitations

• Small Original Dataset Size: The original dataset only contains 35 images, which is a very small number for training a robust image classification model. While data augmentation helps increase the number of examples, it doesn't introduce truly new information or diversity.
• Binary Classification: The dataset is limited to a binary classification problem (two classes). It cannot be used directly for multi-class sign identification without further annotation.
• HEIC Image Handling: The dataset was created by processing HEIC images. While the conversion to PNG was successful, relying on specific image formats and conversion processes can introduce potential issues or dependencies.
• Limited Augmentation Techniques: While several augmentation techniques were used, exploring a wider range of augmentations or more advanced techniques might be necessary for more sensitive applications. Sociotechnical Limitations
• Lack of Diversity: The dataset might lack diversity in terms of variations in lighting conditions, angles, distances, occlusions, or different styles of the same sign, which could impact the model's performance in real-world scenarios. Most importantly, this dataset does not collect images across varying locations or weather conditions, making it unsuitable to be applied broadly.
• Ethical Considerations: Depending on the specific types of signs and their context, there could be ethical implications in developing a sign identification system, such as privacy concerns or the potential for misuse. Though this was carefully mitigated upon initial creation of this dataset, there are still possible safety implications of the widespread availability of these images.
• Generalizability: Due to the small size and potential biases, the dataset is unlikely to be representative of a broad population of signs, limiting the generalizability of models trained on it.

Recommendations

Users should be aware of the risks, biases and limitations of the dataset before use (see above). Refrain from utilizing this dataset for applications outside of algorithm creation optimization/ education. This dataset should not be utilized to draw real-world conclusions.

Dataset Card Authors

Emily Copus

Dataset Card Contact

[email protected]