{"slug": "reclaim-scalable-campus-centric-property-recovery-via-high-dimensional-visual", "title": "ReClaim — Scalable, Campus-Centric Property Recovery via High-Dimensional Visual Search", "summary": "ReClaim, a computer-vision-powered lost-and-found system for campus communities, uses CLIP embeddings and cosine similarity to match lost items with found items. The system, built with FastAPI and PyTorch, enforces a 3:4 crop and allows up to six photos per item, with campus-level isolation to improve search relevance. It targets matches above 90% similarity.", "body_md": "Lost-and-found systems on campuses are difficult to scale. Lost items are usually described with vague text, while found items often sit in administrative offices with little to no structured metadata.\n\n**ReClaim** tackles this problem by turning lost-and-found into a **computer-vision-powered visual search system** built specifically for campus communities.\n\nInstead of relying solely on descriptions like *\"black water bottle lost near the library,\"* users can upload photos of an item and let an AI-powered visual search pipeline identify visually similar items across the campus.\n\nThe goal is simple:\n\nUpload → Search visually → Find a match → Reclaim your property.\n\nReClaim combines a mobile-first reporting experience with a vision inference pipeline and campus-scoped vector search.\n\n```\n┌─────────────────────────┐\n│     Image Capture       │\n│       3:4 Crop          │\n└────────────┬────────────┘\n             │\n             ▼\n┌─────────────────────────┐\n│    FastAPI Backend      │\n│     Python + PyTorch    │\n└────────────┬────────────┘\n             │\n             ▼\n┌─────────────────────────┐\n│    CLIP Vision Model    │\n│   Image → Embedding     │\n└────────────┬────────────┘\n             │\n             ▼\n┌─────────────────────────┐\n│  Cosine Similarity      │\n│   Vector Comparison     │\n└────────────┬────────────┘\n             │\n             ▼\n┌─────────────────────────┐\n│   Match Thresholding    │\n│     > 90% Similarity    │\n└────────────┬────────────┘\n             │\n             ▼\n┌─────────────────────────┐\n│ Firestore + Notification│\n│      Match Found        │\n└─────────────────────────┘\n```\n\nBuilt using:\n\nThe camera workflow enforces a **3:4 aspect-ratio crop**, creating standardized visual inputs before they reach the inference pipeline.\n\nUsers can upload **up to 6 photos per item**, allowing the system to capture different perspectives, textures, colors, and identifying characteristics.\n\nThe backend is powered by:\n\nImages are processed through the CLIP vision encoder to generate dense visual embeddings.\n\nThese embeddings capture high-level visual characteristics such as:\n\nRather than storing only human-written descriptions, ReClaim creates a machine-readable representation of the item's visual identity.\n\nWhen a user reports an item, the uploaded images are normalized through the camera pipeline.\n\nThe standardized **3:4 crop** helps reduce irrelevant background information and creates more consistent model inputs.\n\nUsers can provide multiple images to improve the representation of the object.\n\nEach image is passed through the CLIP vision transformer.\n\n```\nInput Image\n     │\n     ▼\nCLIP Vision Encoder\n     │\n     ▼\nVisual Features\n     │\n     ▼\nDense Embedding Vector\n```\n\nThe resulting embedding represents the visual characteristics of the object in a high-dimensional numerical space.\n\nFor our implementation, the generated representation is treated as a **512-dimensional embedding vector**.\n\nWhen a new lost or found item is submitted, its embedding is compared against active items within the same campus.\n\nThe backend calculates **cosine similarity** between the target embedding and candidate embeddings.\n\n```\nTarget Item\n     │\n     ▼\nEmbedding A ─────────┐\n                     │\n                     ▼\n              Cosine Similarity\n                     ▲\n                     │\nEmbedding B ─────────┘\n     │\n     ▼\nSimilarity Score\n```\n\nHigh-scoring candidates are surfaced as potential matches.\n\nA configurable similarity threshold is used to filter low-confidence results, with the current workflow targeting matches above **90% similarity**.\n\nOne of ReClaim's key design decisions is **campus-level isolation**.\n\nInstead of searching a massive global database, queries are restricted to the user's verified campus.\n\nFor example:\n\n```\nUser\n │\n ▼\nRV University\n │\n ├── Lost Items\n ├── Found Items\n └── Reclaimed Items\n```\n\nThis significantly reduces the search space while making the results more relevant.\n\nA lost laptop at one university should not be compared against thousands of unrelated laptops reported across an entire city.\n\nReClaim uses **Firebase** as its backend platform.\n\nThe system leverages:\n\nThe core Firestore structure is organized around campuses:\n\n```\n/campuses/{campusId}/items/{itemId}\n\n  ├── itemId: String\n  ├── reporterId: String\n  ├── status: LOST | FOUND | RECLAIMED\n  ├── name: String\n  ├── description: String\n  ├── locationFound: String\n  ├── dropoffDestination: String\n  ├── imageURLs: Array[String]\n  └── embeddings: Array[Float]\n```\n\nPotential matches are stored separately:\n\n```\n/campuses/{campusId}/matches/{matchId}\n\n  ├── targetItemId: String\n  ├── candidateItemId: String\n  ├── similarityScore: Float\n  └── timestamp: Timestamp\n```\n\nThis separation allows the system to maintain the original item records while independently tracking generated match relationships.\n\nDespite the complexity of the AI backend, ReClaim is designed around a simple principle:\n\nThe user shouldn't need to understand the AI.\n\nThe interface is optimized for fast reporting and high-visibility interaction.\n\nA verified campus badge such as:\n\n**RV University • Main Campus**\n\nmakes the search boundary explicit to the user.\n\nThe home screen immediately separates the two primary workflows:\n\n**\"Lost Something?\"**\n\nand\n\n**\"Found Something?\"**\n\nThe retro pixel-art cards create a distinctive visual identity while making the application's primary actions immediately recognizable.\n\nUsers can provide:\n\nFound-item reports can also specify a physical destination such as:\n\n**Admin Block Reception**\n\nThis connects the digital discovery system with the physical lost-and-found process.\n\nPotential matches are presented as image cards containing:\n\nThis allows users to visually inspect potential matches instead of manually searching through hundreds of text-based reports.\n\nReClaim is designed around **verified campus communities**.\n\nAuthentication through Firebase allows the platform to associate reports with authenticated users while maintaining campus-level boundaries.\n\nThis creates a more trustworthy environment than a completely open lost-and-found marketplace.\n\nTraditional lost-and-found systems depend heavily on metadata.\n\nFor example:\n\n```\n\"Black bottle\"\n\"Lost near library\"\n\"Found in Block B\"\n```\n\nThe problem is that users rarely describe objects consistently.\n\nTwo users might describe the same object as:\n\n```\n\"Black Hydro Flask\"\n\"Black water bottle\"\n\"Metal bottle\"\n\"Black flask with sticker\"\n```\n\nA visual embedding provides another layer of information.\n\nInstead of asking:\n\n**\"Did someone use the exact same words?\"**\n\nReClaim asks:\n\n**\"Does this object look like the one someone reported?\"**\n\nMany campus objects look extremely similar.\n\nTwo black water bottles may have nearly identical shapes but belong to different people.\n\nBecause of this, similarity scores should be treated as **candidate matches rather than absolute proof**.\n\nPhotos can contain desks, floors, people, walls, and other irrelevant visual information.\n\nThe standardized **3:4 cropping workflow** helps make the object itself more prominent within the model input.\n\nA single photograph may not capture enough information to distinguish an object.\n\nSupporting up to **6 photos per report** allows ReClaim to build a stronger visual representation from multiple viewpoints.\n\nAs the number of reports grows, comparing every new image against every stored embedding becomes increasingly expensive.\n\nThe current architecture establishes the foundation for eventually introducing dedicated vector-search infrastructure and approximate nearest-neighbor indexing for much larger campuses.\n\nReClaim's current architecture can be extended in several directions:\n\nA particularly interesting future direction is **multimodal retrieval**, where both the uploaded image and the user's textual description contribute to the final ranking.\n\nReClaim isn't simply another lost-and-found form.\n\nIt combines:\n\n**Computer Vision + Vector Embeddings + Campus Authentication + Real-World Item Recovery**\n\ninto a single workflow.\n\nThe system connects the entire process:\n\n```\nPhysical Object\n      ↓\nPhoto\n      ↓\nAI Embedding\n      ↓\nCampus-Scoped Search\n      ↓\nPotential Match\n      ↓\nUser Verification\n      ↓\nPhysical Reclamation\n```\n\nThe ultimate goal is to make lost-and-found systems **searchable, scalable, and intelligent**.\n\nReClaim is designed around a simple observation:\n\n**Lost-and-found desks already have the objects. The problem is finding the right person.**\n\nBy turning physical objects into searchable visual representations, ReClaim aims to bridge that gap.\n\nInstead of manually browsing lists of vague descriptions, students can simply show the system what they're looking for.\n\n**See it. Search it. Reclaim it.**\n\n| Layer | Technology |\n|---|---|\n| Mobile | React Native, Expo, TypeScript |\n| Styling | Tailwind CSS |\n| Backend | Python, FastAPI |\n| AI / ML | PyTorch, CLIP |\n| Authentication | Firebase Auth |\n| Database | Firestore |\n| Distribution | Firebase App Distribution |\n| Search | Dense Vector Embeddings + Cosine Similarity |\n\n**A campus-centric visual search engine for lost and found property.**\n\nDon't describe what you lost. Show us.\n\n([https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/hd3romox7ofy2qfde0uv.jpg](https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/hd3romox7ofy2qfde0uv.jpg))\n\n([https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/px74zp1u2l103ivmnta9.jpg](https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/px74zp1u2l103ivmnta9.jpg))\n\n([https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/mqtbyg6f7usvh1qvljwv.jpg](https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/mqtbyg6f7usvh1qvljwv.jpg))\n\n([https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/d72qb7nkd5jlpc8dldzl.jpg](https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/d72qb7nkd5jlpc8dldzl.jpg))\n\n([https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/bd26qcqxr09h34xrsi8s.jpg](https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/bd26qcqxr09h34xrsi8s.jpg))", "url": "https://wpnews.pro/news/reclaim-scalable-campus-centric-property-recovery-via-high-dimensional-visual", "canonical_source": "https://dev.to/neeraj_ciju/reclaim-scalable-campus-centric-property-recovery-via-high-dimensional-visual-search-2106", "published_at": "2026-08-16 07:41:27+00:00", "updated_at": "2026-08-16 08:12:00.094962+00:00", "lang": "en", "topics": ["computer-vision", "machine-learning", "ai-products"], "entities": ["ReClaim", "FastAPI", "PyTorch", "CLIP", "Firestore"], "alternates": {"html": "https://wpnews.pro/news/reclaim-scalable-campus-centric-property-recovery-via-high-dimensional-visual", "markdown": "https://wpnews.pro/news/reclaim-scalable-campus-centric-property-recovery-via-high-dimensional-visual.md", "text": "https://wpnews.pro/news/reclaim-scalable-campus-centric-property-recovery-via-high-dimensional-visual.txt", "jsonld": "https://wpnews.pro/news/reclaim-scalable-campus-centric-property-recovery-via-high-dimensional-visual.jsonld"}}