Datasets:

jzyee
/

SynDORBench

ArXiv:

Tags:

License:

Dataset card Files Files and versions

xet

Community

Acknowledge the license to access SynDORBench

SynDORBench is released for non-commercial research under CC BY-NC 4.0, for evaluation only.

By requesting access you agree to use SynDORBench under the CC BY-NC 4.0 license — for non-commercial purposes only, with attribution — to use it solely for evaluation and benchmarking and NOT to train, fine-tune, or otherwise fit models on it, and to cite the SynDORBench paper in any resulting work.

SynDORBench: Evaluating LVLM Perceptual Robustness Under Physically Constrained Visibility Conditions

Figure 1. Conceptual overview of SynDORBench: The benchmark systematically varies physical visibility factors (lighting, scene type, camera azimuth/elevation, action pose) and DORI-calibrated viewing distance, evaluating models on 3 tasks: 1) binary human-presence classification and 2) action recognition (MCQ and open-ended)

Description

SynDORBench is a synthetic benchmark for robust perception under controlled distance and lighting conditions. Scenes span the DORI regions (Detection / Observation / Recognition) distance scale crossed with lighting bands (day / night / night-vision), for fine-grained evaluation of how perception degrades at range and in low light.

Figure 2. Sample images from SynDORBench.

SyndDORBench consists of 87k QA pairs:

action recognition (open-ended): 27k
action mcq : 27k
human-presence classification : 27k + 5k empty scenes

SynDORBench/
├── images/                 # rendered frames + annotations 
└── index/
    ├── 20k_index.jsonl  # unified record index across all tasks
    ├── human_detection/    # human-presence classification task
    │   ├── {day,night,nv}.jsonl         # GT person-present index
    │   ├── empty_{day,night,nv}.jsonl   # empty-scene (person-absent) negatives
    ├── action_recognition/ # open-ended action recognition (similarity-scored)
    │   └── {day,night,nv}.jsonl         # action-label index
    └── action_mcq/         # 8-way multiple-choice action recognition
        └── {day,night,nv}.jsonl         # MCQ index

Modalities: day = RGB daytime · night = RGB low-light · nv = "night-vision" achieved through image enhancement.

Citation

If you use SynDORBench, please cite the SynDORBench paper.

@article{yee2026syndorbench,
  title={SynDORBench: Evaluating LVLM Perceptual Robustness Under Physically Constrained Visibility Conditions},
  author={Yee, Jeremy Stephen Gabriel and Wang, Zhengkui and Zhang, Zhiyuan and Anand, Avinash and Liu, Timothy and Chan, Benedict and Ng, Aik Beng and See, Simon},
  journal={arXiv preprint arXiv:2609.31823},
  year={2026}
}