0408happyfeet/p3hw1-pen-detection

P3HW1 — Pen Detection (Binary) — 224×224

Purpose

Binary image classification: does the image contain a pen? A compact dataset for course assignments and demos.

Dataset Composition

• —Original split (normalized to 224×224): 34 images
• —pen: 17
• —no pen: 17
• —Augmented split (also 224×224): 306 images
• —Contains the 34 resized originals plus 272 label‑preserving augmentations.

All 34 original images were student‑captured (17 pen, 17 no pen) under varied backgrounds, lighting, and viewpoints.

Data Collection

• —Source: personal, student‑created photos only.
• —Locations: indoor settings (desk/tabletop).
• —Timeframe: 2025-09.
• —No faces or sensitive PII present.

Preprocessing & Augmentation

• —Resolution: all images are 224×224 (both splits).
• —Augmentations (label‑preserving, torchvision):
• —RandomHorizontalFlip(0.5)
• —RandomRotation(±15°)
• —RandomAffine(translate ≤10%, scale 0.9–1.1, shear 10°)
• —ColorJitter(brightness ±0.2, contrast ±0.15, saturation ±0.1, hue ±0.02)
• —RandomAdjustSharpness(×1.5)
• —GaussianBlur(ks=3, σ∈[0.1, 1.5])
• —Excluded: heavy perspective, cutout/erase, aggressive crops that could remove the pen.

Labels

• —label (ClassLabel): 0 → no pen, 1 → pen
• —label_str: string alias

Splits

• —original: resized originals with metadata (orig_width, orig_height, width=224, height=224, source="original").
• —augmented: resized originals + augmented images (source in {original, augmented_v1}), all 224×224.

Features / Metadata Columns

Each example includes:

• —image — 224×224 RGB
• —label (ClassLabel) and label_str (string)
• —orig_width, orig_height (size before normalization)
• —width, height (post‑normalization; always 224)
• —source ("original" or "augmented_v1")

If needed, stratify into train/val/test from the original and mirror indices into augmented.

Intended Use & Limits

• —Teaching/demo for binary presence classification.
• —Not for production or safety‑critical applications.
• —Small object size, occlusion, or very low light may degrade performance.

Ethics & Safety Notes

• —All images are inanimate objects; no biometric data.
• —Backgrounds may include incidental printed text.

License

• —Dataset: CC BY 4.0 (attribution required).

AI Usage Disclosure

• —No generative AI was used to create original images.
• —An LLM (ChatGPT) assisted in writing the collection/augmentation/upload code; images in the augmented split were produced by deterministic torchvision transforms, not by a generative model.