Skip to content
Edition
Late EditionM.S. Computer Science · 3.97 GPA
The Sreekaran Reddy Portfolio
← Back to Projects
Computer Vision Project

PhotoCull

Privacy-First Camera Roll Intelligence

A local-first computer-vision system for detecting redundant photos, organizing events, and building explainable photo shortlists without paid inference APIs.

  • Computer Vision
  • Ranking
  • Clustering
  • Local ML
  • Privacy-First

Try the Web Demo

Upload a small set of images for temporary session analysis, or use the bundled demo. The walkthrough above the demo tells you what to press and what to expect at each step. It is a guide only: it cannot see inside the demo, so you confirm each step yourself.

Step 1 of 12 · Start

0 done

Choose your photos

How do you want to try PhotoCull?
What to do

Pick how you want to start. The instructions for your choice appear here.

If the demo says it is busy or out of resources, click Clear temporary session and try again. An idle app may also take a moment to wake.

What to expect

With the sample, results appear in seconds. With uploads, you add and analyze photos one at a time.

How it works

Both routes run the same local computer-vision pipeline on the demo server. No paid AI API is involved.

What comes next

Once photos are loaded, the walkthrough moves to the results.

Live demoNew tab

Interactive demo

The demo runs best in its own tab on small screens.

Full Local Edition. The complete version scans local folders and persists analysis privately on the user’s machine.

Hosted on Streamlit Community Cloud. An idle app may take a moment to wake. Open the full demo in a new tab.

Pipeline

From camera roll to shortlist

Each stage is cheap, local, and explainable on its own. Optional embedding models add signal when they are installed; when they are not, the pipeline degrades gracefully to the hash-based path.

  1. Photo scanDecode-check each file, record metadata and a SHA-256 fingerprint. Originals are never modified.
  2. Technical-quality featuresMeasurable quality indicators, computed after EXIF orientation is applied.
  3. Perceptual hashesCompact visual fingerprints that stay close when two images look alike.
  4. Duplicate / burst detectionExact matches by SHA-256, near-duplicates by a conservative hash rule, bursts by capture time.
  5. Event groupingLikely occasions from capture-time gaps, with uncertain timestamps marked low confidence.
  6. RankingExplainable, diversity-aware shortlists; each duplicate or burst group contributes at most one pick.
  7. Optional personalizationOptionalExplicit pairwise choices train a local preference model.

Evaluation

Measured before it was trusted

Synthetic held-out benchmark

98.36%Near-duplicate precision
71.43%Near-duplicate recall
82.76%Near-duplicate F1
155Automated tests passed
0.54s1K-image warm run

Duplicate metrics come from a synthetic held-out benchmark, not real camera rolls. They are not a real-world accuracy claim, and preferences have not been validated with people.

Runtime and validation results
CheckResult
1,000-image local fixture run13.31s cold / 0.54s warm
30-image web analysis core0.33s
Automated tests155 passed
Known dependency vulnerabilities0 at final validation
Engineering takeaway

Embedding hybrids improved recall on standard pairs but introduced more hard-negative false positives, so the conservative hash-based system remained the production default.

Engineering

What makes it technically interesting

  • SHA-256 + perceptual hashing

    Exact and visual fingerprints, side by side.

  • Conservative near-duplicate rule

    Tuned to avoid flagging distinct photos.

  • Timestamp-aware bursts

    A burst marks a moment, not redundancy.

  • Event clustering

    Capture-time grouping with manual corrections that persist.

  • Local visual embeddings

    Optional MobileNet / TinyCLIP, run on-device.

  • Diversity-aware ranking

    Shortlists that cover the set instead of repeating it.

  • Pairwise preference learning

    Human-in-the-loop choices, kept local.

  • Evaluation-first model selection

    Defaults chosen on held-out and hard-negative sets.

  • Offline, private inference

    No paid inference API; degrades gracefully without models.

Privacy

Two editions, two privacy models

Local Edition

Photos stay on-device

  • Scans folders on your machine
  • Local SQLite cache
  • Persistent analysis and decisions
  • Photos never leave the device

Web Demo

Session-only processing

  • Explicit browser uploads
  • Processed temporarily for the session
  • No persistent user database
  • Images do reach the demo server

What PhotoCull never does automatically

  • Delete, move or rename photos
  • Edit your photographs
  • Upload Local Edition photos to a server
  • Identify people with face recognition
  • Make permanent aesthetic judgments

It is a decision-support tool: you decide what stays and what goes. Web Demo uploads are processed temporarily for the session and are not intentionally persisted; only the Local Edition keeps photos on the device. No paid inference API is used in either edition.