Abstract
Machine Learning (ML) solutions are rapidly evolving and are increasingly capable of performing Automated Visual Data Processing (AVDP) tasks such as visual scene understanding, 3D model reconstruction, and automated content generation. We believe that novel AVDP solutions can significantly boost a company's business, streamline the creation of 3D content and models, enable adaptation of content in XR applications, and support the creation of digital twins. However, there are also obstacles that limit the business use of such solutions. The objective of our research is to study the skills and insights of companies in using different kinds of AVDP solutions in their business. This article presents the results of interviews with 10 Finnish companies. The interviews comprised three sections: a brief introduction to existing AVDP solutions, a background questionnaire on ML experience and interests, and thematic interviews covering specific visual data processing themes of interest to the companies.
Introduction: Bridging the Gap Between Research and Industry
While the academic world is flooded with papers on computer vision and generative AI, a significant disconnect remains regarding how these technologies actually "boost business" in real-world industrial settings. Companies face a steep climb when trying to integrate tasks like 3D model reconstruction, autonomous scene understanding, or automated content generation into their workflows.
This study identifies the specific skills, interests, and obstacles Finnish companies face. By interviewing representatives from various sectors, the researchers aimed to determine what companies actually need from AVDP, the roles they expect to play in the ecosystem (developer vs. user), and the primary bottlenecks preventing them from moving from experimentation to tangible business value.
Methodology
The researchers conducted a qualitative study consisting of semi-structured interviews with 17 representatives from 10 Finnish companies, ranging from micro-enterprises to large corporations in the manufacturing and service sectors. The study was structured in three phases:
- State-of-the-Art Briefing: Researchers provided a classification of current ML-based AVDP solutions (Scene Understanding, 3D Reconstruction, and Content Generation) to ensure a common vocabulary.
- Background Questionnaire: Collected data on the companies' current ML experience and future interests.
- Thematic Interviews: Deep dives into specific visual data processing themes selected by the companies as relevant to their business.
The "Data Desert" Paradox
As a qualitative study, the research did not train new models but instead uncovered a critical "dataset gap" within the industry. This gap is characterized by a paradox: while 80% of the interviewed companies have raw, unlabelled data available for training, only 10% actually possess the labeled data necessary for production-grade ML. Furthermore, only one company reported using simulated training data, highlighting an underutilized avenue for overcoming data scarcity.
Industrial Insights and Strategic Shifts
The interviews revealed several key insights into how businesses view the future of AVDP:
- The Rise of the "Hybrid" Company: The traditional divide between "tech providers" and "end users" is blurring. 70% of the companies expect to be both developers and users of AVDP solutions in the future.
- Reliability Over Breadth: In an industrial context, companies prefer a system that performs one task with 98% accuracy over a system that attempts ten things with 10% accuracy. For professional use, such as building design, the tolerance for error is near zero.
- The "Boring" Problems are the Hardest: Companies are not just looking for high-end generative models; they need tools for the "manual" labor of ML—automated labeling, cleaning up noisy point clouds, and simplifying complex 3D models for mobile use.
- UX as a Competitive Edge: Especially in Extended Reality (XR), the "wow" effect and seamless User Experience (UX) are seen as just as important as the underlying ML accuracy.
Key Results and Adoption Trends
- High Adoption Interest: 90% of the companies had already experimented with ML, though only 30% had successfully integrated it into their core business operations.
- Demand for 3D: There is massive interest in 3D model reconstruction from point clouds (70%) and the synthesis of indoor scenes (80%). Companies see these as keys to creating "digital twins" and streamlining construction workflows.
- Scene Understanding: 80% of companies identified a need for basic object detection, but 70% are already looking toward more complex tasks like 3D object pose estimation.
- Cloud Efficiency: One of the primary drivers for AVDP interest is the proven cost-efficiency of current cloud-based ML solutions.
Conclusion and Future Directions
The study’s primary limitation is its sample size of 10 Finnish companies with pre-existing interests in AVDP, which may not represent the broader global market. However, the findings suggest clear paths for future work:
- Developing better tools for automated data labeling to help companies bridge the "data desert."
- Researching automated 3D model simplification to allow high-fidelity reconstructions to run on low-power mobile devices.
- Exploring simulated data generation as a standard business practice to train models when real-world labeled data is too expensive to produce.
Summary updated