106,977 grayscale image crops packaged for training classical face detectors like Viola-Jones and Haar cascades. The dataset includes a large pool of natural-image negatives from Caltech-256 and a specific CBCL benchmark split for testing. It was uploaded by user salvacarrion to Hugging Face and last updated on 2026-05-14.
Use Cases
- Training sliding-window classifiers based on the provided grayscale face crops and negative images.
- Benchmarking Viola-Jones or Haar cascade detectors using the CBCL benchmark test split.
- Augmenting training data with the large pool of 29,879 natural-image negatives from Caltech-256.
Strengths
- The training split contains 102,429 face images, providing a substantial positive sample.
- Includes a dedicated test split of 24,045 images with a known CBCL benchmark.
- Offers a large, separate pool of 29,879 negative images from the Caltech-256 dataset for robust classifier training.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- The dataset consists solely of grayscale images, which may limit applicability for color-based detection methods.
- Row count for the overall dataset is unknown, which may limit suitability assessment.
Provenance
- Source
- huggingface
- Collection Method
- Packaged from existing sources including a CBCL benchmark and Caltech-256 negatives.
- Freshness
- Last updated 2026-05-14 21:59:14; freshness should be verified.