IDMR: Towards Instance-Driven Precise Visual Correspondence in Multimodal Retrieval
π arXiv 2504.00954 | π» GitHub
The IDMR-test dataset is designed to evaluate instance-driven multimodal retrieval.
We provide two types of evaluation scenarios:
In-domain evaluation: derived from benchmarks also used in training (COCO, Objects365, OpenImages).
Out-of-domain⦠See the full description on the dataset page:
https://huggingface.co/datasets/lbw18601752667/IDMR-test.