MVRB Leaderboard
MVRB (Massive Visualized IR Benchmark) evaluates multimodal retrieversβ performance on general Vis-IR tasks. The benchmark includes various task types, such as screenshot-based multimodal retrieval (screenshot to anything, anything to screenshot) and screenshotconditioned retrieval (e.g., searching for documents using queries conditioned on screenshots). It also covers a variety of important domains, including news, products, papers, and charts.
More details can be found:
- "headers": [
- "Rank",
- "Model",
- "#Params (B)",
- "Overall",
- "SR",
- "CSR",
- "SQA",
- "OVC"
- "data": [
- [
- 1,
- "<a href="https://huggingface.co/BAAI/BGE-VL-Screenshot">BGE-VL-Screenshot</a>",
- 3.75,
- 60.61,
- 70.09,
- 59.58,
- 53.1,
- 54.46
- [
- 2,
- "<a href="https://huggingface.co/Alibaba-NLP/gme-Qwen2-VL-2B-Instruct">GME</a>",
- 2.21,
- 48.14,
- 61.62,
- 37.68,
- 37.78,
- 47.98
- [
- 3,
- "<a href="https://huggingface.co/Tevatron/dse-phi3-v1.0">DSE</a>",
- 4.15,
- 45.21,
- 61.54,
- 37.78,
- 39.24,
- 31.51
- [
- 4,
- "<a href="https://huggingface.co/vidore/colpali">ColPali</a>",
- 2.92,
- 43.64,
- 61.73,
- 35,
- 35.32,
- 31.04
- [
- 5,
- "<a href="https://huggingface.co/nvidia/MM-Embed">MM-Embed</a>",
- 7.57,
- 34.48,
- 25.86,
- 40.93,
- 42.83,
- 32.67
- [
- 6,
- "<a href="https://huggingface.co/google/siglip-so400m-patch14-384">SigLIP</a>",
- 0.878,
- 33.34,
- 38.33,
- 34.48,
- 19.6,
- 40.64
- [
- 7,
- "<a href="https://huggingface.co/TIGER-Lab/VLM2Vec-Full">VLM2Vec</a>",
- 4.15,
- 32.19,
- 15.93,
- 48.05,
- 49.42,
- 23.24
- [
- 8,
- "<a href="https://huggingface.co/royokong/e5-v">E5-V</a>",
- 8.35,
- 25.13,
- 34.11,
- 26.59,
- 5.23,
- 32.85
- [
- 9,
- "<a href="https://huggingface.co/openai/clip-vit-large-patch14">CLIP</a>",
- 0.428,
- 23.75,
- 18.89,
- 25.39,
- 23.9,
- 30.4
- [
- 10,
- "<a href="https://huggingface.co/TIGER-Lab/UniIR">Uni-IR</a>",
- 0.428,
- 19.63,
- 12.35,
- 35.92,
- 29.68,
- 20.06
- [
- 11,
- "<a href="https://huggingface.co/OpenDriveLab/Vista">VISTA</a>",
- 0.196,
- 13.85,
- 5.21,
- 11.29,
- 25.78,
- 16.61
- [
- "metadata": null
Tasks
Screenshot Retrieval (SR) consists of evaluation samples, each comprising a textual query q and its relevant screenshot s: (q, s). The retrieval model needs to precisely retrieve the relevant screenshot for a testing query from a given corpus S. Each evaluation sample is created in two steps: 1) sample a screenshot s, 2) prompt the LLM to generate a search query based on the caption of screenshot. We consider seven tasks under this category, including product retrieval, paper retrieval, repo retrieval, news retrieval, chart retrieval, document retrieval, and slide retrieval.
Composed Screenshot Retrieval (CSR) is made up of sq2s triplets. Given a screenshot s1 and a query q conditioned on s1, the retrieval model needs to retrieve the relevant screenshot s2 from the corpus S. We define four tasks for this category, including product discovery, news-to-Wiki, knowledge relation, and Wiki-to-product. All tasks in this category are created by human annotators. For each task, annotators are instructed to identify relevant screenshot pairs and write queries to retrieve s2 based on s1.
Screenshot Question Answering (SQA) comprises sq2a triplets. Given a screenshot s and a question q conditioned on s, the retrieval model needs to retrieve the correct answer a from a candidate corpus A. Each evaluation sample is created in three steps: 1) sample a screenshot s. 2) prompt the MLLM to generate a question q. 3) prompt the MLLM to generate the answer a for q based on s. The following tasks are included in this category: product-QA, news-QA, Wiki-QA, paper-QA, repo-QA.
Open-Vocab Classification (OVC) is performed using evaluation samples of screenshots and their textual class labels. Given a screenshot s and the label class C, the retrieval model needs to discriminate the correct label c from C based on the embedding similarity. We include the following tasks in this category: product classification, news-topic classification, academic-field classification, knowledge classification. For each task, we employ human labelers to create the label class and assign each screenshot with its correct label.
Make sure you submit your evaluation results in a JSON file with the following format:
{
"Model": "<Model Name>",
"URL (optional)": "<Model/Repo/Paper URL>"
"#params": "7.11B",
"Overall": 30.00,
"SR": 30.00,
"CSR": 30.00,
"VQA": 30.00,
"OVC": 30.00,
}
Then send a email to lz123@mail.ustc.edu.cn with the JSON file attached. We will review your submission and add it to the leaderboard.