A due diligence assessment evaluating model architectures, dataset availability, deployment constraints, and ROI timelines for AI-powered office furniture detection on mobile devices.
This project delivered a due diligence assessment for Digital Xalience AS, evaluating the technical feasibility of building an AI-powered system to detect office furniture in images and video. The work surveyed state-of-the-art computer vision models, available datasets, and deployment frameworks to determine whether a production-ready solution could be built with current technology, and under what constraints.
Objectives
To guide the project from concept toward implementation without compromising user experience or privacy standards, the analysis focused on three core pillars: model performance, data availability, and deployment practicality.
The goal was to:
- Identify model architectures suitable for real-time furniture detection on mobile and edge devices.
- Assess the readiness and gaps of existing datasets for office-specific furniture categories.
- Compare deployment options (on-device, hybrid edge-cloud, cloud-only) in terms of performance, cost, and compliance.
- Provide actionable, evidence-based recommendations for moving forward.
Key Findings
The investigation revealed a landscape where hardware capabilities have caught up with algorithmic needs for many standard tasks, though specific gaps remain when targeting modern office furniture categories. The findings balance performance metrics against practical constraints such as energy efficiency and regulatory compliance. Below is a high-level synthesis of the current state-of-the-art models, dataset limitations, and deployment options identified during this review:
- Model readiness: Lightweight detection models (YOLOv11, YOLOv12) and segmentation models (EdgeSAM) can achieve >30 FPS on modern smartphones with acceptable accuracy. Open-vocabulary models (YOLO-World, Grounding DINO) enable detection of novel furniture types without retraining.
- Dataset gaps: Public datasets cover generic furniture well but lack sufficient examples of modern office items (standing desks, ergonomic chairs, smart furniture). Office-specific categories represent less than 15% of annotated instances in major benchmarks.
- Deployment maturity: Frameworks like TensorFlow Lite/LiteRT, Core ML, and ONNX Runtime provide mature tooling for mobile optimization. Hybrid edge-cloud architectures offer a practical balance between performance and cost.
- Cost and compliance: On-device deployment eliminates per-inference fees but requires $60–130 per device in hardware. Cloud-only inference costs $1–5 per thousand images. GDPR and licensing restrictions (e.g., non-commercial datasets) demand careful data handling and model selection.
Conclusions
Ultimately, the assessment confirms that immediate development is viable using current technology stacks. However, success depends on addressing dataset limitations regarding modern workplace environments and adhering to strict privacy regulations like GDPR before scaling operations. A phased approach—starting with well-represented furniture categories and expanding through targeted data collection—is recommended for risk mitigation.
This page provides a high-level summary. For technical specifications, performance tables, and implementation roadmaps, refer to the complete Technical Feasibility Report (to be released).