COCO-OD-TR-Single-Objects-v2: Turkish-Labeled Object Detection Images

A combined and deduplicated version of the COCO-2014 and COCO-2017 datasets for object detection. The dataset features labels in Turkish and is formatted for instruction-tuning with separate prompt and completion columns. It was created by user 'ucsahin' and last updated on August 6, 2024.

Use Cases

• Fine-tune vision-language models for object detection based on Turkish-language prompts.
• Train instruction-following models on a multimodal dataset based on the described instruction-tuning format.
• Benchmark object detection models on a Turkish-labeled version of the COCO dataset.
• Study cross-lingual transfer learning for computer vision tasks based on the Turkish annotations.

Strengths

• Combines and deduplicates two established datasets: COCO-2014 and COCO-2017.
• Provides Turkish-language labels, which are less common than English annotations for this task.
• Uses a structured annotation scheme for bounding boxes, similar to PaliGemma, with normalized coordinates.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect the geographic or cultural bias inherent to the original COCO datasets.

Provenance

Source

Combined from COCO-2014 and COCO-2017 datasets.

Collection Method

Combined and deduplicated by the author.

Freshness

Last updated 2024-08-06 16:50:37; freshness should be verified.