BengaliFoodSeg: Traditional Bengali Food Images for Detection and Segmentation

1.4 GB of high-quality images of traditional Bengali dishes captured in diverse real-life settings. The dataset is annotated with polygon-based masks and categorized into multiple food classes, supporting semantic segmentation and object detection tasks. Created by Ahmed Imtiaz and last updated in May 2026, it was used to train UNet and YOLOv12 models.

Use Cases

• Train food detection models based on polygon-based mask annotations.
• Benchmark semantic segmentation performance on culturally specific food items.
• Develop dietary assessment tools based on annotated images of traditional dishes.
• Improve model generalization for food recognition using described augmentation and class balancing techniques.
• Research culturally contextualized computer vision based on the dataset's focus on Bengali cuisine.

Strengths

• Dataset size is 1.4 GB, indicating a medium-scale collection of images.
• Annotations include polygon-based masks, providing detailed segmentation ground truth.
• Exports are available in both COCO and mask formats, offering flexibility for different model architectures.
• Augmentation and class balancing techniques were applied, which likely improves model generalization.

Limitations

• Row count is unknown, which may limit suitability assessment for large-scale training.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect geographic or cultural bias inherent to its focus on traditional Bengali dishes.

Provenance

Source

figshare

Collection Method

Images were annotated and preprocessed using the Roboflow platform.

Freshness

Last updated 2026-05-31 03:22:38.

Geography

Likely contains images from regions where traditional Bengali cuisine is prevalent.