Pinpoint-Bench: A Zero-Hint Evaluation Frontier for Active Visual Search

Project Page | Paper (PixelEyes)

Pinpoint-Bench is a specialized evaluation benchmark designed to assess the active visual search and reasoning capabilities of Multimodal Large Language Models (MLLMs) and active perception agents within ultra-high-resolution images. It explicitly addresses the saturation of traditional benchmarks, the lack of spatial annotations, and the vulnerability of models to text-guided shortcuts.

Dataset Summary

• Total Samples: 433 meticulously human-annotated samples.
• Image Resolution: Ultra-high-resolution images, with an average resolution of 5500x3516.
• Target ROI Ratio: Target objects are extremely tiny, accounting for an average of only 0.07% of the total image area (a genuine "needle-in-a-haystack" challenge).
• Annotation Types: Each target object is simultaneously annotated with both detailed Bounding Boxes and Instance Masks, enabling comprehensive root-cause analysis of multi-turn navigation errors (e.g., distinguishing between perception bottlenecks and reasoning failures).

Key Features

1. Strict Zero-Hint Protocol: All spatial location priors (e.g., "in the bottom left corner," "on the huge bridge") are entirely removed from the questions. Models must independently rely on autonomous, exhaustive visual search and regional zooming to uncover the critical evidence, which ranges from fine-grained property recognition and micro-OCR to spatial relationship reasoning.

2. Diagnostic Evaluation Metrics:

   • LSR (Localization Success Rate): Decoupled from the final textual answer accuracy, a search trajectory is deemed successful if the model's multi-turn cropping history covers the ground-truth target mask at the pixel level. The discrepancy between LSR and Accuracy directly quantifies the "Inattentional Blindness" phenomenon—where a model successfully locates the target region but fails to correctly output the final answer due to weak reasoning or long-context degradation.
   • TAE (Turn-to-Answer Efficiency): Balances model performance against interactive costs, formulated as $\text{TAE} = \frac{\text{Accuracy}}{\text{AvgTurns}}$. Higher TAE signifies that an agent can accurately pinpoint answers using fewer interactive turns.

3. Robust Multi-Alias Scoring: To accommodate natural language ambiguities arising from extremely subtle visual attributes (e.g., describing a "beige" backpack as "white" or "light brown"), each question is equipped with an exhaustive set of alternative answer aliases. This is combined with an LLM-as-a-Judge grading pipeline to ensure robust and fair scoring.

Citation

If you find this dataset helpful in your research, please cite our paper:

@misc{gong2026pixeleyesdecouplingperceptionreasoning,
      title={PixelEyes: Decoupling Perception and Reasoning for Pinpoint Visual Evidence Seeking}, 
      author={Dengxian Gong and Yuanzheng Wu and Haobo Yuan and Zhengdong Hu and Tao Zhang and Yikang Zhou and Shihao Chen and Quanzhu Niu and Kai Wang and Jason Li and Haochen Wang and Lu Qi and Shunping Ji and Ming-Hsuan Yang},
      year={2026},
      eprint={2607.00115},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.00115}, 
}