Tony2202/Paper_pythera

Dataset Gender

Dataset Gender - gender classification (female/male/unknown) for crowded scenes, YOLO format.

1. Data sources & preprocessing

• Filtered from CrowdHuman (~5,000 images), removing severely misaligned/overlapping boxes.
• Augmented with Pexels & Unsplash images using keywords: asian_multiple_kids, japanese_kids, japanese_multiple_kids, elderly, grandfather, grandmother (~250 images per keyword).
• Deduplicated; prioritized images with clear faces/upper bodies for reliable gender cues.

2. Labels & format

• Classes: 0=female, 1=male, 2=unknown.
• YOLO label files: <class> <x_center> <y_center> <width> <height> normalized to [0,1] relative to image size (body bounding boxes).
• Original head/full_body tags are discarded; only body boxes suitable for gender recognition are kept.
• data.yaml currently points to sample absolute paths; update to your local images/ and labels/ locations.

3. Scale & splits

• Total curated images: 6,382.
• Splits: train 5,896 images, val 236 images, test 250 images (1-1 alignment between images and labels).
• Label distribution (approx): train – female ~9k, male ~8k, unknown ~2.2k; val/test keep similar ratios, unknown is the smallest class.
• Boxes are predominantly tall/vertical; object centers cluster near the image center; objects per image mostly 1–10, very few >15.

4. Directory layout

images/
  train/ 5896 images
  val/    236 images
  test/   250 images
labels/
  train/ 5896 txt
  val/    236 txt
  test/   250 txt
visualize/ (statistical plots)
data.yaml (paths + class names)

5. Visualization & analysis

5.1. Aspect Ratio (w/h)

Analysis

• Train: Most boxes cluster near ratio 0; peak ~7,500 at ~0, then sharp drop. Majority in 0–2; very few >4.
• Val: Peak near 0 (~340). Most in 0–1.5; few >2.
• Test: Peak ~160 in 0.0–0.1. Most in 0–0.5; very few >1.5.
• Observation: Boxes are largely vertical or near-square, consistent with human posture (standing/sitting).

5.2. Bounding-box area (normalized)

Analysis

• Train: Peak ~7,500 near area 0, then sharp decay. Most areas 0–0.2; rare >0.6.
• Val: Peak ~820 near area 0; similar shape to train. Most 0–0.4; rare >0.6.
• Test: Peak ~400 near area 0; gradual decay. Most 0–0.4; very few >0.8 at low frequency.
• Observation: Boxes usually cover a small fraction of the image, implying either small object footprint or high image resolution; pattern is consistent across splits.

5.3. Objects per image

Analysis

• Train: Highest frequency (~58 images) has 6 objects; ~40 images have 7 objects. Counts 5–10 range from 20–60 images. Very few images (>15 objects) — fewer than 5.
• Val: Highest (~1,850 images) has 1 object; a few hundred images have 5–10. Counts drop as objects increase; very few >15.
• Test: Highest (~135 images) has 1 object; ~40 images have 5–10. Counts drop with more objects; very few >15.
• Observation: Non-uniform object-count distribution; most images have 1–6 objects, tapering quickly. Models should handle sparse-object cases well while remaining robust to rare crowded cases.

5.4. Label distribution

Analysis

• Train: Male ~8k, female ~9k, unknown ~2.2k.
• Val: Male ~950, female ~700, unknown 200–300.
• Test: Male ~450, female ~450, unknown ~100.
• Observation: Male slightly dominates overall; unknown is the smallest class, reflecting ambiguous cases. Distribution is stable across splits.

5.5. Bounding-box center heatmap

Analysis

• Train: Highest density (~200) at the image center; gradual decay outward.
• Val: Highest (~9) at center; more even spread than train but lower overall density.
• Test: Highest (~17.5) at center; more concentrated than val, with rapid decay outward.
• Observation: Objects are primarily centered, likely due to framing or ROI selection. Train shows highest density due to larger sample size; val/test follow the same pattern with lower counts.