almador2002/Cars196

<div align="center">

🚗 Cars196 — Visual Recommendation System

Find visually similar cars from an image or a text description

    

A CLIP-powered visual recommender that finds the 3 most similar cars to any input — using 512-dimensional joint image-text embeddings and cosine similarity.

</div>

🎯 What This Does

Upload a car photo or type a natural-language description — the app returns the 3 most visually similar cars from a dataset of ~8,000 vehicles spanning 196 fine-grained classes.

📹 Presentation Video

<div align="center">

<!-- Replace YOURVIDEOID with the actual YouTube video ID after uploading --> 

Click the thumbnail above to watch the full 5-minute walkthrough on YouTube

</div>

🧠 Tech Stack at a Glance

<div align="center">

</div>

🗺️ The Full Journey

The complete analytical pipeline lives in Assignment_3_Cars196.ipynb. Here are the highlights.

📊 Part 1 — Dataset Exploration

Cars196 is a fine-grained classification benchmark containing 16,185 images of 196 car classes, where each class corresponds to a specific Make + Model + Year combination (e.g., "BMW 3 Series Sedan 2012"). We use the training split (~8,144 images).

[image]

🔍 Part 2 — Exploratory Data Analysis

A deep dive into the dataset's structure — class balance, manufacturer distribution, body types, production years, and image dimensions.

[image]

Key findings

• —📦 Roughly balanced — ~40 images per class on average
• —🏭 American and European manufacturers dominate the dataset
• —🚙 Sedans, coupes, and SUVs are the most represented body types
• —📅 Cars range from 1991 to 2012 — with a heavy skew toward 2007-2012

[image]

[image]

🧬 Part 3 — Embeddings & Clustering

All ~8,000 car images were passed through CLIP ViT-B/32, producing L2-normalized 512-dimensional vectors.

[image]

3.1 — Dimensionality Reduction

Compared PCA, t-SNE, and UMAP side by side:

[image]

UMAP produced the most semantically meaningful clusters — visually similar cars (same body type, similar proportions) end up close together regardless of brand.

[image]

3.2 — Clustering

Applied two complementary clustering algorithms:

[image]

[image]

[image]

3.3 — Cluster Interpretation

CLIP embeddings cluster cars primarily by:

1. 1.🚗 Body type / silhouette — the strongest signal (SUVs together, sedans together, sports cars together)
2. 2.🎨 Visual style & proportions — low-slung sports cars separate from tall boxy SUVs
3. 3.💡 Color & lighting tendencies — backgrounds & palettes influence proximity
4. 4.📏 Size category — compact cars, mid-size sedans, and full-size vehicles form distinct groups

[image]

Verdict — are the clusters useful? Yes. They capture the visual feel of a car rather than just its brand name. A user looking for a sporty coupe naturally finds similar cars across BMW, Audi, and Porsche — exactly what a recommender should do.

🎯 Part 4 — The Recommendation Engine

User Input (image or text)
        ↓
    CLIP encoder
        ↓
   512-D vector (L2-normalized)
        ↓
   Cosine similarity vs. all 8,144 stored embeddings
        ↓
        ↓  ← single matrix multiplication
        ↓
    Top-3 ranked matches

Image Search Test

[image]

Text Search Tests

CLIP's killer feature: text and images live in the same vector space, so a phrase like "red sports car" can match images directly with no extra training.

[image]

🎁 Bonus Analyses

Beyond the assignment's core requirements, the notebook explores a few extras:

1. Body Type Similarity Heatmap

How similar are different body types in CLIP's vector space? Sedans and coupes are close; trucks and convertibles are far apart.

[image]

2. Embedding Interpolation

Linearly interpolating between a Sedan and a Convertible in the embedding space — and watching the nearest car at each step morph from one to the other.

[image]

3. Nearest Neighbor Exploration

For a random query car, what are its closest neighbors in embedding space?

[image]

⚖️ Ethical Considerations

A few things to keep in mind about this system:

• —Dataset bias — Cars196 is heavily skewed toward American and European brands. Asian manufacturers (Toyota, Honda, Hyundai) are underrepresented, which biases recommendations toward Western models.
• —Temporal bias — Most cars in the dataset are from 2007–2012. The system won't recognize modern (2020+) designs well.
• —Environmental angle — A more responsible recommender could highlight fuel efficiency or electric alternatives. This system is purely visual — it doesn't consider environmental impact.
• —Privacy — Uploaded images are processed in-memory and never stored. In a production setting, users should be explicitly informed about data handling.

📂 Repository Contents

cars196-recommender/
├── app.py                            # The Gradio application
├── requirements.txt                  # Python dependencies (pinned for HF Spaces compatibility)
├── README.md                         # ← You are here
├── car_embeddings.parquet            # Pre-computed 512-D embeddings + metadata
├── car_images.pkl                    # Compressed thumbnails for display
└── Assignment_3_Cars196.ipynb        # Full notebook (EDA → embeddings → clustering → recommendations)

🛠️ How to Run Locally

git clone https://huggingface.co/spaces/YOUR_USERNAME/cars196-recommender
cd cars196-recommender
pip install -r requirements.txt
python app.py

Then open http://localhost:7860 in your browser.

📚 Acknowledgments & References

• —CLIP — Radford et al., Learning Transferable Visual Models From Natural Language Supervision (paper)
• —Cars196 / Stanford Cars — Krause et al., 3D Object Representations for Fine-Grained Categorization (paper)
• —UMAP — McInnes et al., Uniform Manifold Approximation and Projection (paper)
• —HuggingFace Datasets — for distributing Cars196 in a clean, accessible format
• —Gradio — for making ML demos trivially easy to share

<div align="center">

Built for Assignment #3 — RecSys, Embeddings & Spaces

🤗 Hosted on HuggingFace Spaces • 🧠 Powered by CLIP • 💜 Made with Gradio

</div>