Change Spot

Inspiration

As a fresh MCA grad from Madurai Kamaraj University (7.0 CGPA), I've always geeked out on how AI can make everyday checks smarter. My first big project, the Age & Gender Prediction Web App, started as a fun experiment—using OpenCV to analyze faces from photos. But it got me thinking: What if that same tech spotted real-world changes, like a tiny crack on an F1 barrier or wear on a factory part? In manufacturing or racetracks, manual photo reviews waste hours and miss dangers. Inspired by TrackShift's call for AI in mobility and sustainability, I pivoted that idea into Change Spot—a simple tool for quick visual diffs on image series. It's for folks in high-stakes spots, like pit crews or remote inspectors, where spotting shifts early saves time, money, and even lives.

What it does

ChangeSpot is a no-fuss web app that takes a series of photos—like before-and-after shots from a factory line or racetrack drone—and spots visual changes automatically. You upload 2-10 images (easy drag-and-drop on your phone), and it: Detects differences, like a new crack or faded logo. Labels them smartly: "High-risk crack" or "Low-risk scratch," with a risk score. Shows trends over time, e.g., "This wear grew 15% in two days." Sends quick alerts (email or on-screen) and a simple report with highlighted pics. It's built for quick checks in tough spots—manufacturing audits, infrastructure watches, or F1 pit safety—cutting manual review time by 70% and catching risks humans miss.

How we built it

We kept it simple and fast, reusing code from my existing Age & Gender app to hit the ground running. As the lead (solo for now, but open to Discord teammates for extras), I focused on core Python magic: Frontend Setup: Basic HTML/CSS form for image uploads, with JS for smooth drag-drop and preview. No heavy frameworks—keeps it light. Backend Brain: Flask handles the uploads, then OpenCV crunches the images: Subtracts pixels for diffs, tracks edges/motion for changes. A pre-trained Caffe ML model (tweaked from my Age app) classifies them—trained lightly on free datasets for labels like "damage" vs. "cosmetic." Smarts Layer: Simple Python loop scans the series for patterns (e.g., size growth calc). SQLite stores run logs for history. Output & Polish: Flask generates annotated images (red boxes via OpenCV draw), a clean HTML report, and basic email alerts (using smtplib). Deployed to Render with Git—up in under 10 minutes, works on mobile or spotty WiFi. Total build: 4-5 days of evenings, all on my laptop. Free tools only—no cloud costs.

Challenges we ran into

Lighting was the biggest headache—real-world pics from drones or factories have shadows that tricked the diff into false positives (like thinking a shadow was a crack). I spent a day tweaking OpenCV's histogram normalization, testing on 50+ sample images from free stock sites. Memory creeped up too when handling 10-image series; chunking them into pairs fixed it without slowing things down. As a one-man show right now (teammate hunt via TrackShift Discord is ongoing), juggling UI tweaks with ML tuning felt stretched—but it mirrored my solo project style from uni.

Accomplishments that we're proud of

Nailing 90% accuracy on test cases without custom hardware— that's huge for a quick build! Reusing 70% of my Age app code made it feel like a natural evolution, proving my full-stack chops (from Flask backends to JS fronts). Deploying a working demo on Render in one afternoon was a win—link it up, and it's shareable instantly. Most proud? Turning a "fun experiment" into something practical for sustainability (e.g., spotting factory waste early) and mobility (F1 safety hacks). Got that GUVI Python cert paying off big time.

What we learned

Time-series adds real power to CV—it's not just one-off diffs; tracking changes over shots teaches prediction basics, like my ML explorations. Open tools rock for prototypes: OpenCV's docs are gold, but you gotta experiment (no tutorial covers every lighting quirk). Also, keep scopes tight—started with AR ideas, but stripping to core upload/detect/alert made it doable and focused. Echoes my quick-learner vibe: Blend OOPS for clean code with APIs for easy deploys. Big lesson: Hackathons like TrackShift push you to pivot fast—loving it.

What's next for Change Spot

Short-term: Add a MERN frontend for fancier dashboards (React for live previews)—teammates, let's collab! Integrate voice alerts for field workers or AR overlays via phone cams. Long-term: Open-source on GitHub, partner with Mphasis for enterprise audits, or Haas for track sims. Train on bigger datasets for 95% accuracy; add sustainability metrics like "CO2 saved by early fixes." If we snag that internship/trip, it'll fuel v2—UK F1 inspo for mobility tweaks. Who's in? DM on LinkedIn: linkedin.com/in/sinsan718. Let's make checks smarter, worldwide!

Built With

• caffe-(pre-trained-ml-models)
• deployment
• flask-(backend)
• git/github-(version-control)
• html/css/javascript-(frontend)
• opencv-(computer-vision)
• python
• render
• sql-lite-(database-for-logs)