Evaluation of YOLOv8 and Faster R-CNN for Food-Scene Detection on the Central Asian Food Scene Dataset

Abukhassym Khydyrbayev, Zhanar Oralbekova, Elvira Aitmukhanbetova, Makpal Zhartybayeva

2026

Data confidence 70 medium

How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

This research investigates deep learning–based object detection architectures for food-scene recognition using the Central Asian Food Scene Dataset (CAFSD). We compare a one-stage detector (YOLOv8) and a two-stage detector (Faster RCNN) using standard object detection metrics, including mAP@50, precision, recall, and F1-score. CAFSD contains 239 food categories with 17,046 training images, 2,084 validation images, and 2,176 test images, representing a challenging benchmark due to high intra-class variation and inter-class similarity. Using transfer learning from COCO-pretrained models, YOLOv8 achieved mAP@50 ≈ 0.68, precision ≈ 0.72, and recall ≈ 0.65, while Faster R-CNN achieved mAP@50 ≈ 0.74 and recall ≈ 0.69 under the same evaluation setting. The results highlight the classic speed–accuracy trade-off: YOLOv8 provides efficient inference suitable for real-time applications, whereas Faster R-CNN offers higher detection accuracy at the cost of increased computational complexity. Overall, this study expands food-scene recognition research for under-represented cuisines and provides reproducible baseline results on CAFSD.