This repo contains evaluation code for the paper "BLINK: Multimodal Large Language Models Can See but Not Perceive" [ECCV 2024]
🌐 Homepage | 🤗 Dataset | 📑 Paper | 💻 Code | 📖 arXiv | 🔗 EvalAI
🔥[2024-07-03]: We added a validation set leaderboard on the website.
🔥[2024-07-01]: BLINK is accepted to ECCV 2024!
🔥[2024-06-16]: We added the scores and outputs for GPT-4o and GPT-4 Turbo.
🔥[2024-05-18]: Gemini 1.5 Pro achieved great improvements on BLINK!
🔥[2024-04-24]: Evaluation for test set is now availble on EvalAI. We welcome all submissions and look forward to your participation!
We introduce BLINK, a new benchmark for multimodal language models (LLMs) that focuses on core visual perception abilities not found in other evaluations. Most of the BLINK tasks can be solved by humans “within a blink” (e.g., relative depth estimation, visual correspondence, forensics detection, and multi-view reasoning). However, we find these perception-demanding tasks cast significant challenges for current multimodal LLMs because they resist mediation through natural language. BLINK reformats 14 classic computer vision tasks into 3,807 multiple-choice questions, paired with single or multiple images and visual prompting. While humans get 95.70% accuracy on average, BLINK is surprisingly challenging for existing multimodal LLMs: even the best-performing GPT-4V and Gemini achieve accuracies of 51.26% and 45.72%, only 13.17% and 7.63% higher than random guessing, indicating that such perception abilities have not “emerged” yet in recent multimodal LLMs. Our analysis also highlights that specialist CV models could solve these problems much better, suggesting potential pathways for future improvements. We believe BLINK will stimulate the community to help multimodal LLMs catch up with human-level visual perception.
BLINK is created to challenge multimodal models on hollistic visual perception abilities with tasks inherited from classic computer vision problems, stimulating future development of multimodal LLMs that achieve human-level visual perception. Please refer to our huggingface 🤗 Dataset for more details. Unique Features of BLINK includes diverse visual prompting, beyond recognition perception, and visual commonsense.
import datasets
dataset_name = 'BLINK-Benchmark/BLINK'
data = load_dataset(dataset_name, SUBTASK_NAME)
where SUBTASK_NAME is one of the subtasks: ['Art_Style', 'Functional_Correspondence', 'Multi-view_Reasoning', 'Relative_Reflectance', 'Visual_Correspondence', 'Counting', 'IQ_Test', 'Object_Localization', 'Semantic_Correspondence', 'Visual_Similarity', 'Forensic_Detection', 'Jigsaw', 'Relative_Depth', 'Spatial_Relation']
Please refer to our eval folder for more details.
| Model | Val (1,901) | Test (1,907) |
|---|---|---|
| 🏅 Human | 95.7 | 95.7 |
| 🥈 GPT-4o | 60.0 | 59.0 |
| 🥉 GPT-4 Turbo | 54.6 | 53.9 |
| GPT-4 (vision preview) | 51.1 | 51.3 |
| Gemini Pro 1.0 | 45.2 | 45.7 |
| LLaVA-1.6-34B | 46.8 | 45.1 |
| Claude 3 Opus | 44.1 | 44.1 |
| Yi-VL-34B | 41.7 | 42.8 |
| Qwen-VL-MAX | 40.3 | 42.0 |
| LLaVA-v1.5-13B-xtuner | 42.0 | 41.3 |
| Yi-VL-6B | 38.7 | 41.7 |
| LLaVA-v1.5-7B-xtuner | 39.4 | 40.8 |
| LLaVA-1.5-13B | 42.7 | 40.6 |
| InstructBLIP-13B | 42.2 | 39.6 |
| CogVLM | 41.5 | 39.4 |
| InstructBLIP-7B | 39.7 | 38.7 |
| OpenFlamingo2-9B | 39.2 | 38.3 |
| 👀 Random Choice | 38.1 | 38.1 |
| LLaVA-1.5-7B | 37.1 | 38.0 |
| LLaVA-internLM2-7B | 37.7 | 36.1 |
| MiniGPT-4-v2-7B | 34.2 | 34.6 |
🎯 We have released a full suite comprising 1,901 validation samples, the prompt we used, and model predictions for the baselines tested in our paper. However, the 1,907 test questions are available without their answers. You can submit your model's predictions for the test set on EvalAI.
Blink makes use of data from existing image datasets, and does not cover all the visual perception abilities in the wild. For the forensics detection task, we manually collected images that are publicly available from online search. We have made every effort to ensure that the images included in this paper are used in accordance with applicable copyright laws and are properly credited. However, if you are the copyright owner of any image included in our work and believe that its use conflicts with your licensing agreements, please contact us directly. We are committed to addressing any legitimate concerns promptly.
- Xingyu Fu: xingyuf2@seas.upenn.edu
- Yushi Hu: yushihu@uw.edu
- Wei-Chiu Ma: weichiu@cornell.edu
- Ranjay Krishna: ranjay@cs.washington.edu
BibTeX:
@article{fu2024blink,
title={BLINK: Multimodal Large Language Models Can See but Not Perceive},
author={Fu, Xingyu and Hu, Yushi and Li, Bangzheng and Feng, Yu and Wang, Haoyu and Lin, Xudong and Roth, Dan and Smith, Noah A and Ma, Wei-Chiu and Krishna, Ranjay},
journal={arXiv preprint arXiv:2404.12390},
year={2024}
}