Technical Summary: TimeLens

Problem Statement

The paper addresses three critical challenges in deploying AI-powered mobile guides for the Grand Egyptian Museum (GEM):

1. Fine-Grained Visual Similarity: The target collection includes 51 catalogued artifacts, many of which are near-identical Ramesside statues differing only in subtle pose, headdress, or accompanying standards.
2. Data-Reality Gap: There is a significant discrepancy between curated training data and the conditions of handheld camera capture (low light, background clutter, visual overlap).
3. Hallucination Risk: Standard language models often invent plausible-sounding but historically unsupported facts, which is unacceptable for a cultural heritage guide.

Existing solutions fail to simultaneously meet the requirements of real-time on-device recognition (avoiding cloud latency), native Arabic language support, and hallucination-free accuracy.

Methodology

TimeLens is a bilingual (English/Arabic) mobile application built with Flutter, integrating two primary subsystems: an on-device artifact detector and a Retrieval-Augmented Generation (RAG) guide.

1. On-Device Artifact Detection

The detection pipeline utilizes a YOLOv8n model exported to TensorFlow Lite (FP16). The development followed a rigorous data-quality-driven iteration study:

• v1 (Auto-Annotation): Initial labels were generated using YOLO-World (foundation model) with text prompts. This resulted in poor performance (mAP@0.5 = 0.751) due to mislocalized boxes and sibling-class confusion.
• v2 (Spatial Sanity Rules): Four geometric rules were applied to filter YOLO-World outputs (e.g., removing boxes in the bottom quarter of the frame, selecting the largest area box). This improved mAP@0.5 to 0.867 but failed to resolve specific class confusions.
• v3 (Full Hand Annotation): The final dataset (24,384 images) was manually annotated in Roboflow, focusing on full-artifact bounding boxes and distinguishing similar artifacts. This step was decisive, achieving zero "broken" classes.

The final model is a 5.97 MB asset that runs asynchronously on mid-range Android phones, utilizing a confidence filter (0.75 globally, 0.90 for the visually similar Statue of Siptah) to ensure reliability.

2. Bilingual RAG Guide

The Q&A subsystem grounds responses in a curated ChromaDB knowledge base containing 108 records (54 unique artifacts) documented in both English and Arabic.

• Architecture: The system uses FastAPI to route queries to a locally served Gemma 4 E2B (Q4 K M) model via Ollama.
• Grounding Strategy:
  • Language Separation: English and Arabic chunks are stored in separate collections; the system routes queries based on script detection (>30% Arabic characters).
  • Chunking: Records are split into identity, description, and significance chunks.
  • Retrieval Logic: Camera mode triggers deterministic metadata lookups based on the detected artifact's title. Text mode uses a keyword-based classifier to route queries to specific fields (Location, Period, Material) or similarity search.
• Optimizations: Ten targeted optimizations (including hierarchical chunking, Flash Attention, and adjusted token limits) reduced end-to-end latency from >30s to ~10s.

Key Contributions

The paper reports two primary engineering contributions:

1. Data-Quality-Driven Detection: The study isolates label quality as the decisive factor for fine-grained museum artifact detection. The authors demonstrate that label cleaning and hand annotation yielded greater performance gains (+0.116 and +0.128 mAP@0.5 respectively) than architectural changes or resolution scaling. The final model resolves all previously failing classes.
2. Local, Grounded Bilingual RAG: The system achieves hallucination-free answers in both English and Arabic by strictly constraining the LLM to retrieved context. The authors show that grounding mechanisms (chunking, language separation, query classification) are more effective at preventing fabrication than simply increasing model size.

Results

• Detection Performance: The final YOLOv8n model achieved mAP@0.5 = 0.995 and mAP@0.5:0.95 = 0.924 on the validation set. It runs in approximately 800ms per frame (1.25 FPS) on a mid-tier phone, with no misclassifications observed during field testing across all 51 classes.
• RAG Performance: On a 30-question evaluation set covering text, camera, and out-of-dataset queries, the system provided correct, grounded answers with zero hallucinations and zero truncations.
• Latency: End-to-end response time was reduced to approximately 10.2 seconds (average 5.9s per answer in evaluation), making it interactive for museum visitors.
• Resource Efficiency: The entire system runs on commodity hardware: a mid-range mobile phone for detection and a single small GPU (NVIDIA RTX A5000) for the RAG backend.

Significance and Claims

The paper claims that TimeLens successfully bridges the gap between heavy cloud-based AI services and simplistic QR-code systems. Its significance lies in demonstrating that:

• Label quality dominates architecture: For fine-grained, domain-specific detection, manual verification of labels is more critical than model complexity or input resolution.
• Grounding prevents hallucination: A small, quantized model (Gemma 4 E2B) combined with rigorous retrieval constraints can outperform larger models in producing factual, bilingual historical narratives.
• Feasibility of Local Deployment: A high-accuracy, bilingual museum guide can be deployed on commodity hardware without relying on continuous cloud connectivity, addressing the specific needs of the Grand Egyptian Museum and similar cultural heritage sites.

The authors conclude that while the system is production-ready for the specific 51 artifacts and 108 records tested, future work involves expanding the knowledge base and replacing the keyword-based classifier with a learned intent model.