SUPERGLASSES

This repository contains the dataset for the paper SUPERGLASSES: Benchmarking Vision Language Models as Intelligent Agents for AI Smart Glasses, which has been accepted by CVPR 2026 (Findings).

SUPERGLASSES is the first comprehensive Visual Question Answering (VQA) benchmark built on real-world data entirely collected by smart glasses devices. It comprises 2,422 egocentric image-question pairs spanning 14 image domains and 8 query categories, enriched with full search trajectories and reasoning annotations.

The benchmark is specifically designed to evaluate Vision Language Models (VLMs) in realistic smart glasses usage scenarios, where identifying an object of interest is a critical prerequisite for external knowledge retrieval.

If you want to evaluate your models on this benchmark, please refer to the code in SuperGlasses GitHub repository.

Dataset Structure

The dataset is provided in two main configurations:

• images: Contains the egocentric images captured by smart glasses.
• queries: Contains questions, answers, and detailed annotations including difficulty, location, and reasoning sub-questions.

Benchmark Usage

Setup

# clone the evaluation code
git clone https://github.com/xandery-geek/SuperGlasses.git
cd SuperGlasses

# create a conda environment
conda create -n superglasses python=3.10
conda activate superglasses
pip install -r requirements.txt

Loading the Dataset

from datasets import load_dataset

queries = load_dataset("xandery/SuperGlasses", name="queries", split="test")
images  = load_dataset("xandery/SuperGlasses", name="images",  split="test")

Inference

Run inference with any Supported VLM via vLLM:

python main.py \
    --model qwen-7b \
    --hf_repo_name xandery/SuperGlasses \
    --output_dir output \
    --batch_size 8

If you want to add a custom model, please refer to the implementation in our code and append the model to mllm/model.yaml:

my-model:
  model_path: "org/model-name"
  max_model_len: 8192
  # optional: prompt_template, stop_tokens, hf_overrides, ...

Evaluation

Evaluation uses two complementary metrics:

• F1-Recall — token-level recall against the ground-truth answer
• LLM Judge Accuracy — an LLM scores each prediction as correct or incorrect

# Evaluate by fields (hops, dynamism, domain, category, difficulty, language, tools)
python -m evaluation.llm_judger --input output/results_xxx.jsonl

# Evaluate overall scores only
python -m evaluation.llm_judger --input output/results_xxx.jsonl --overall

# Skip LLM inference and score a pre-evaluated file
python -m evaluation.llm_judger --input output/results_xxx.jsonl --from_file

By default the judge model is Qwen/Qwen2.5-32B-Instruct. Override with --model:

python -m evaluation.llm_judger \
    --input output/results_xxx.jsonl \
    --model meta-llama/Llama-3.1-8B-Instruct

Citation

If you use SUPERGLASSES in your research, please cite:

@inproceedings{jiang2026superglasses,
  title={SUPERGLASSES: Benchmarking Vision Language Models as Intelligent Agents for AI Smart Glasses},
  author={Jiang, Zhuohang and Yuan, Xu and Qu, Haohao and Lin, Shanru and Liu, Kanglong and Fan, Wenqi and Qing, Li},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={2165--2175},
  year={2026}
}