Mindykkyan/PittadsDB-AdsPics

Pitt Ads Dataset (PittAdsDB)

This folder holds the University of Pittsburgh Ads dataset artifacts we can download publicly, and a helper script to fetch images once access is granted.

What is included (downloaded now)

• docs/
  • readme_images.txt (dataset readme for images)
  • readme_videos.txt (dataset readme for videos)
• image_annotations/
  • image/QA_Action.json
  • image/QA_Combined_Action_Reason.json
  • image/QA_Reason.json
  • image/Sentiments.json
  • image/Sentiments_List.txt
  • image/Slogans.json
  • image/Strategies.json
  • image/Strategies_List.txt
  • image/Symbols.json
  • image/Topics.json
  • image/Topics_List.txt
  • readme_images.txt
  • annotations_images.zip (original archive)
• video_annotations/
  • video/cleaned_result/*.json
  • video/raw_result/*.json
  • video/final_video_id_list.csv
  • video/Sentiments_List.txt, video/Topics_List.txt
  • readme_videos.txt
  • annotations_videos.zip (original archive)
• video_urls/
  • final_video_id_list.csv (list of video IDs/URLs)
• images/ (populated with 64,832 images organized in subfolders 0-10)
• metadata/ (JSON files and CSV manifests for categorization)

Image Collection Status

✅ DOWNLOADED: 64,832 image ads successfully obtained

The images are organized in the following structure:

• images/0/ - 5,673 images
• images/1/ - 5,433 images
• images/2/ - 5,599 images
• images/3/ - 5,694 images
• images/4/ - 5,716 images
• images/5/ - 5,677 images
• images/6/ - 5,616 images
• images/7/ - 5,657 images
• images/8/ - 5,620 images
• images/9/ - 5,799 images
• images/10/ - 8,348 images

Total: 64,832 images in JPEG format

Metadata and Categorization

The metadata/ directory contains comprehensive categorization data:

• category_topic_sentiment.json - Maps product categories to image files with associated emotional sentiments
• ads_maj_topic_two_sents.json - Major topic classifications with sentiment pairs
• pitt_image_manifest.csv - Complete image manifest with metadata
• pitt_image_local_manifest.csv - Local file mapping

How to use categorization data:

1. By Category: Use category_topic_sentiment.json to find images by product category (e.g., "clothing", "cars", "food", "beauty")

2. By Sentiment: Each category contains sentiment-based groupings like:

   • Emotional states: "cheerful", "confident", "inspired", "calm"
   • Personality traits: "fashionable", "educated", "active", "creative"
   • Attitudes: "persuaded", "amazed", "eager", "youthful"

3. By Topic: Access topic classifications through the annotation files

Example usage:

import json
with open('metadata/category_topic_sentiment.json') as f:
    data = json.load(f)
    
# Get all "cheerful" clothing ads
cheerful_clothing = data["clothing"]["cheerful"]
# Returns list of image filenames: ["7160.jpg", "90950.jpg", ...]

# Get all beauty-related images for "fashionable" sentiment
fashionable_beauty = data["beauty"]["fashionable"] 
# Returns: ["19782.jpg", "53570.jpg", "130250.jpg", ...]

Working with Downloaded Images

The images are ready for use! Here are some ways to work with the dataset:

Finding specific images:

# Count images in each category
ls -la images/*/

# Find a specific image (e.g., from sentiment analysis)
find images/ -name "19782.jpg"  # Returns: images/2/19782.jpg

# List all images from a specific subfolder
ls images/5/ | head -10

Using with BrainDive Analysis:

# Example: Create analysis manifest for clothing ads with "cheerful" sentiment
import json, os

with open('metadata/category_topic_sentiment.json') as f:
    data = json.load(f)

cheerful_clothing = data["clothing"]["cheerful"]
image_paths = []

for filename in cheerful_clothing:
    # Find the image in the appropriate subfolder
    for i in range(11):  # subfolders 0-10
        path = f"images/{i}/{filename}"
        if os.path.exists(path):
            image_paths.append(path)
            break

print(f"Found {len(image_paths)} cheerful clothing ads for analysis")

(Optional) Download videos

• The file video_urls/final_video_id_list.csv contains video IDs/URLs.
• You can use a tool like yt-dlp to download videos if needed.

Example (optional):

# Install yt-dlp (optional)
pip install yt-dlp

# Download a subset of videos
python - <<'PY'
import csv, subprocess, os
root = '/data/mindy/PittAdsDB'
os.makedirs(os.path.join(root, 'videos'), exist_ok=True)
with open(os.path.join(root, 'video_urls', 'final_video_id_list.csv')) as f:
    r = csv.DictReader(f)
    for i, row in enumerate(r):
        if i >= 10: break  # demo: first 10
        url = row.get('url') or row.get('video_url') or row.get('link')
        if not url: continue
        subprocess.run(['yt-dlp', '-P', os.path.join(root,'videos'), url])
PY

Integrating with BrainDive

• You can map downloaded image filenames to our existing dior_roi_activations_data.csv-style schema to analyze ROI activations for 3rd-party ads.
• We can create a converter once you choose a subset of images to evaluate.

Provenance

• Source page: https://people.cs.pitt.edu/~kovashka/ads/#image
• Image annotations: https://people.cs.pitt.edu/~kovashka/ads/annotations_images.zip
• Video annotations: https://people.cs.pitt.edu/~kovashka/ads/annotations_videos.zip
• Readmes: readme_images.txt, readme_videos.txt

If you want me to request access or batch-download given a URL list, share the file and I’ll run the downloader.