MarketVision, computer vision for market monitoring

Traditional wet markets remain essential for food security in the Global South, yet their informal
nature leaves food flows invisible to researchers and policymakers. This paper presents
MarketVision, a probe testing whether computer vision can generate reliable, low-cost data on
informal markets from smartphone imagery. We developed YOLOv8-based models for object
detection (mAP50 = 0.72) and segmentation (mAP50 = 0.60) trained on a dataset of 1,419 images
covering ten common vegetables and fruits found in Vietnamese wet markets. The models were
validated in a simulated market environment using multiple smartphone devices under varying
conditions. Additionally, we demonstrated that photogrammetry can produce coherent 3D
reconstructions from walkthrough videos, supporting spatial localization and reducing double
counting. Results show that computer vision offers a scalable approach for monitoring food
diversity and availability in settings that currently lack reliable data. The method’s reliance on
standard smartphones and open-source tools makes it accessible for national statistical offices,
municipalities, and research partners in low-resource contexts. Future work will focus on
operationalizing data collection protocols, expanding the product taxonomy, and integrating
automatic price extraction to enable comprehensive market monitoring.