TDBench: Benchmarking Vision-Language Models in Understanding Top-Down Images

Kaiyuan Hou*, Minghui Zhao*, Lilin Xu, Yuang Fan, Xiaofan Jiang (* equally contributing first authors)

Intelligent and Connected Systems Lab (ICSL), Columbia University · NeurIPS 2026 (Evaluations & Datasets Track)

Paper | Code

TDBench evaluates vision-language models on top-down (drone-altitude) images: 2,000 questions in 10 categories, each asked under four image rotations (RotationalEval), plus four case studies (zoom-in, altitude, object integrity, depth).

Usage

The benchmark is run with the TDBench code, a fork of VLMEvalKit. Images from ERA and the Semantic Drone Dataset are not stored here (rows without an image in the viewer); tdbench build creates them from the official downloads of the two datasets:

git clone https://github.com/Columbia-ICSL/TDBench.git && cd TDBench && pip install -e .
tdbench build --era /path/to/ERA_Dataset.zip --sdd /path/to/semantic_drone_dataset_semantics_v1.1.zip
tdbench run main --model Qwen2.5-VL-7B-Instruct

Files

File	Content
tdbench/main.tsv	9 multiple-choice categories, 1,800 questions × 4 rotations
tdbench/grounding.tsv	visual grounding, 200 questions × 4 rotations
tdbench/case_study_{zoom_in,integrity,height,depth}.tsv	case studies
tdbench/rebuild.jsonl	crops and processing steps for the ERA and Semantic Drone images
data/*.parquet	the same tables for the dataset viewer

License

TDBench's questions, annotations and rebuild recipes are released under CC BY-NC-SA 4.0. Images remain subject to the terms of their sources:

Source	License / terms	Images
Aerial Traffic Images	MIT	hosted
VisDrone	CC BY-NC-SA 3.0	hosted
AFO	CC BY-NC-SA 3.0 IGO	hosted
SeaDronesSee	CC0 1.0	hosted
Semantic Drone Dataset	non-commercial academic use, no redistribution	not hosted, tdbench build
ERA	academic use only, no redistribution on other public websites	not hosted, tdbench build
CARLA	code MIT, assets CC BY 4.0; images rendered by the authors	hosted
GTA V	proprietary game; images rendered by the authors	hosted

Citation

@article{hou2025tdbench,
  title={TDBench: Benchmarking Vision-Language Models in Understanding Top-Down Images},
  author={Hou, Kaiyuan and Zhao, Minghui and Xu, Lilin and Fan, Yuang and Jiang, Xiaofan},
  journal={arXiv preprint arXiv:2504.03748},
  year={2025}
}