README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- image-text-to-text
tags:
- vqa
- vision-language-model
- top-down-images
- aerial-images
- benchmark
configs:
- config_name: default
  data_files:
  - split: main_0deg
    path: data/main_0deg-*
  - split: main_90deg
    path: data/main_90deg-*
  - split: main_180deg
    path: data/main_180deg-*
  - split: main_270deg
    path: data/main_270deg-*
  - split: case_study_zoom_in
    path: data/case_study_zoom_in-*
  - split: case_study_integrity
    path: data/case_study_integrity-*
  - split: case_study_height
    path: data/case_study_height-*
  - split: case_study_depth
    path: data/case_study_depth-*
dataset_info:
  features:
  - name: index
    dtype: int64
  - name: image
    dtype: image
  - name: question
    dtype: string
  - name: A
    dtype: string
  - name: B
    dtype: string
  - name: C
    dtype: string
  - name: D
    dtype: string
  - name: answer
    dtype: string
  - name: category
    dtype: string
  splits:
  - name: main_0deg
    num_bytes: 81665845.0
    num_examples: 1800
  - name: main_90deg
    num_bytes: 81873677.0
    num_examples: 1800
  - name: main_180deg
    num_bytes: 81664996.0
    num_examples: 1800
  - name: main_270deg
    num_bytes: 81872635.0
    num_examples: 1800
  - name: case_study_zoom_in
    num_bytes: 29567696.0
    num_examples: 930
  - name: case_study_height
    num_bytes: 49904448.5
    num_examples: 1500
  - name: case_study_integrity
    num_bytes: 48365566.0
    num_examples: 1344
  - name: case_study_depth
    num_bytes: 14894886.0
    num_examples: 400
  download_size: 629990525
  dataset_size: 469809749.5
---

# TDBench: Benchmarking Vision-Language Models in Understanding Top-Down / Bird's Eye View Images

[Kaiyuan Hou](https://hou-kaiyuan.github.io/)+, [Minghui Zhao](https://scottz.net/)+, [Lilin Xu](https://initxu.github.io/), [Yuang Fan](https://www.linkedin.com/in/yuang-fan/), [Xiaofan Jiang](http://fredjiang.com/) (+: Equally contributing first authors)

#### **Intelligent and Connected Systems Lab (ICSL), Columbia University**

[Paper](https://huggingface.co/papers/2504.03748) | [Code / Project Page](https://github.com/Columbia-ICSL/TDBench)

<p align="center">
   <img src="images/TDBench.jpg" width="600"></a>
</p>
<p align="center"> 8 Representative VLMs on 10 dimensions in TDBench </p>

**Abstract:** Top-down images play an important role in safety-critical settings such as autonomous navigation and aerial surveillance, where they provide holistic spatial information that front-view images cannot capture. Despite this, Vision Language Models (VLMs) are mostly trained and evaluated on front-view benchmarks, leaving their performance in the top-down setting poorly understood. Existing evaluations also overlook a unique property of top-down images: their physical meaning is preserved under rotation. In addition, conventional accuracy metrics can be misleading, since they are often inflated by hallucinations or "lucky guesses", which obscures a model's true reliability and its grounding in visual evidence. To address these issues, we introduce TDBench, a benchmark for top-down image understanding that includes 2000 curated questions for each rotation. We further propose RotationalEval (RE), which measures whether models provide consistent answers across four rotated views of the same scene, and we develop a reliability framework that separates genuine knowledge from chance. Finally, we conduct four case studies targeting underexplored real-world challenges. By combining rigorous evaluation with reliability metrics, TDBench not only benchmarks VLMs in top-down perception but also provides a new perspective on trustworthiness, guiding the development of more robust and grounded AI systems. Project homepage: this https URL


## 📢 Latest Updates
- **Apr-23-25**: Submitted [pull request](https://github.com/open-compass/VLMEvalKit/pull/947) to VLMEvalKit repository.
- **Apr-10-25**: Arxiv Preprint is released [arxiv link](https://arxiv.org/abs/2504.03748). 🔥🔥
- **Apr-01-25**: We release the benchmark [dataset](https://huggingface.co/datasets/Columbia-ICSL/TDBench).
---

## 💡 Overview
<p align="center">
   <img src="images/dimension_examples.jpg" width="1200"></a>
</p>

## 🏆 Contributions

- **TDBench Benchmark.** We introduce TDBench, a benchmark designed specifically for evaluating VLMs on Top-down images originate from real scenarios is aerial operation or drone applications. We carefully curated a dataset manually comprising a total of 2000 questions.
- **Rotational Evaluation.** We introduce an evaluation strategy *RotationalEval* specifically designed for top-down images. Due to the nature of top-down images, rotations do not affect the semantic meaning, whereas this is not true and does not physically make sense naturally for front-view images.
- **Four Case Studies.** We performed 4 case studies that frequently occur in the real world.These studies evaluate specific capabilities of VLMs under controlled conditions, providing actionable insights for practical deployment while identifying critical challenges that must be addressed for reliable aerial image understanding.
<hr />

## 📊 Benchmarks Comparison

<p align="center">
   <img src="images/overall_performance.jpg" width="1200" alt="Dataset Comparison table"></a>
</p>


<p align="center"> Overview performance of 8 open source VLMs and 6 propriety VLMs on 10 dimensions with RotationalEval method. </p>

<hr />


## 🗂️ Case Studies

Top-down images are usually captured from a relatively high altitude, which may introduce several challenges such as small object, different perspective. Furthermore, top-down images do not contain depth information in most cases, yet depth is very important for many real-world applications such as building height estimation and autonomous drone navigation and obstacle avoidance. Based on these considerations, we also conduct the following four case studies in paper.
1. **Digital Magnification for Small Object Detection**
   - Provide insights on post-processing the images to enable VLMs to see small objects

2. **Altitude Effects on Object Detection**
   - Guidelines on drones' hovering height for different object detection tasks

3. **Object Visibility and Partial Occlusion**
   - Study when objects are partially hidden or occluded by other objects

4. **Z-Axis Perception and Depth Understanding**
   - Assessing the depth reasoning from top-down images

## 🤖 Sample Usage: How to run TDBench

TDBench is fully compatible with [VLMEvalKit](https://github.com/open-compass/VLMEvalKit). 

### Installation
1. First, install the VLMEvalKit environment by following the instructions in the [official repository](https://github.com/open-compass/VLMEvalKit)
2. Set up your model configuration and APIs according to VLMEvalKit requirements

### Datasets (for VLMEvalKit run.py)
* **Standard Evaluation** - Tests 9 dimensions with 4 rotation angles
  * `tdbench_rot0` (0° rotation)
  * `tdbench_rot90` (90° rotation)
  * `tdbench_rot180` (180° rotation)
  * `tdbench_rot270` (270° rotation)

* **Visual Grounding** - Tests visual grounding with 4 rotation angles
  * `tdbench_grounding_rot0` (0° rotation)
  * `tdbench_grounding_rot90` (90° rotation)
  * `tdbench_grounding_rot180` (180° rotation)
  * `tdbench_grounding_rot270` (270° rotation)

* **Case Studies** - 4 studies
  * `tdbench_cs_zoom`
  * `tdbench_cs_height` 
  * `tdbench_cs_integrity` 
  * `tdbench_cs_depth` 

### Usage Examples

#### Standard Evaluation
To only evaluate a single rotation
```python
python run.py --data tdbench_rot0 \
              --model <model_name> \
              --verbose \
              --work-dir <results_directory>
```
To apply RotationalEval, simply run all rotations
```python
python run.py --data tdbench_rot0 tdbench_rot90 tdbench_rot180 tdbench_rot270 \
              --model <model_name> \
              --verbose \
              --work-dir <results_directory>
```


#### Visual Grounding Evaluation
To only evaluate a single rotation
```python
python run.py --data tdbench_grounding_rot0 \
              --model <model_name> \
              --verbose \
              --judge centroid \
              --work-dir <results_directory>
```
To apply RotationalEval, simply run all rotations
```python
python run.py --data tdbench_grounding_rot0 tdbench_grounding_rot90 tdbench_grounding_rot180 tdbench_grounding_rot270 \
              --model <model_name> \
              --verbose \
              --judge centroid \
              --work-dir <results_directory>
```

#### Case Studies

Run all case studies with:

```python
python run.py --data tdbench_cs_zoom tdbench_cs_height tdbench_cs_integrity tdbench_cs_depth \
              --model <model_name> \
              --verbose \
              --work-dir <results_directory>
```

### Output
VLMEvalKit prints and saves each dataset's output in `<results_directory>/<model_name>`. Check `xxx_acc.csv` for accuracy score, and `xxx_result.xlsx` for detailed VLM outputs.
RotationalEval is triggered automatically after running all rotations. Results will be printed and saved as `xxx_REresult.csv`.
<hr />

## 📜 Citation
If you find our work and this repository useful, please consider giving our repo a star and citing our paper as follows:
```bibtex
@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}
}
```


## 📨 Contact
If you have any questions, please create an issue on this repository or contact at kh3119@columbia.edu or
mz2866@columbia.edu.

---
[<img src="images/ICSL_Logo.png" width="500"/>](http://icsl.ee.columbia.edu/)