Turkish Image-Captioning Benchmark on MS COCO 2014

Models that understand an image and describe it in a sentence need large, clean data to train on. For English, MS COCO has done that job for years - the field's shared measuring stick. For Turkish there was no equivalent.
The consequence: Turkish vision-language work could not be compared. Everyone built their own small dataset, and results had no common ground to be measured on.
Filling the gap by machine translation alone wouldn't do it either. Automatic translation produces captions that drift from the image or read badly in Turkish, and a benchmark built on those is a faulty measuring stick. What was needed was verified quality at scale.

As part of the research group at Istanbul Bilgi University, we turned MS COCO 2014 into a complete benchmark for Turkish image captioning research.

1. Scale and human verification together. The dataset contains 616,767 Turkish captions covering 123,287 images, and the captions were human-verified - not just machine-translation output. Large enough for serious model training, reliable enough to serve as a measuring stick.
2. We tested the dataset with models. To claim a benchmark works, you have to train on it. Five image-captioning models were trained and compared on the data, ranging from classical CNN+LSTM architectures with a ResNet backbone to the Meshed-Memory Transformer.
3. Results surpassed the state of the art at the time. The trained models exceeded the best prior results in Turkish image captioning, demonstrating both the dataset's quality and the value of comparing architectures on it.
4. Released openly. The dataset was published for the research community, so subsequent work can be measured on the same ground.

The work was published at ICECCME 2022. Co-authors: Sina Berk Golech, Saltuk Buğra Karacan, Elena Battini Sönmez and Hakan Ayral.

Stack: Python, PyTorch; CNN+LSTM (ResNet backbone), Meshed-Memory Transformer and other vision-language architectures; standard captioning evaluation metrics.