PUST Cafeteria Food Image Dataset: Real-World Bangladeshi Meal-Platter Images with Bounding-Box and Polygon Annotations

The PUST Cafeteria Food Image Dataset is a real-world image resource developed for food recognition, object detection, and instance segmentation of Bangladeshi cafeteria meals. The dataset contains 720 original meal-platter photographs collected between 1 August and 5 September 2025 at the Central Cafeteria of Pabna University of Science and Technology, Bangladesh. Images were captured using the rear cameras of Realme GT Master Edition and OPPO A92 smartphones under realistic cafeteria conditions with variations in illumination, viewpoint, background, food arrangement, overlapping objects, and partial occlusions. Each visually distinguishable food instance was manually annotated using the Roboflow platform. The dataset provides bounding-box annotations for object detection and polygon-based annotations for instance segmentation. Annotation preparation involved two annotators with cross-checking for class assignment, bounding-box placement, polygon delineation, and image–annotation consistency. The dataset contains 16 food classes. Images were auto-oriented using available metadata and resized to 640 × 640 pixels. The original images were divided into training, validation, and test subsets using a 70:15:15 split ratio (504, 108, and 108 images). Augmentation and light class balancing were applied only to the training subset, including selective augmentation of five underrepresented classes: chaa, matha, ruti, singara, and vegetable_roll. The final processed dataset contains 1,728 images with 6,297 annotated food instances, including 1,512 training images (5,493 instances), 108 validation images (395 instances), and 108 test images (409 instances). The dataset is provided in three annotation formats: YOLO_detect (bounding-box TXT), YOLO_seg (polygon-based TXT), and COCO (JSON annotation files). The archive includes raw images, processed datasets, metadata files (class_names.csv, dataset_summary.csv, image_source_mapping_final.csv, instances_before_augmentation.csv, instances_after_augmentation.csv, and device_information.txt), together with README.md and LICENSE.txt files. This dataset supports research in food recognition, object detection, instance segmentation, food-instance counting, transfer learning, domain adaptation, benchmarking, and intelligent food-service systems. Applications involving billing, portion estimation, calorie estimation, or dietary assessment require additional validated external information. The dataset was collected from a single university cafeteria using two smartphone models. Class distributions remain non-uniform despite selective augmentation and light class balancing. Augmented images improve variability but cannot replace independently collected meal scenes. Polygon annotations represent only visible food regions and do not reconstruct hidden areas caused by occlusion.