README.md

---
license: apache-2.0
task_categories:
- visual-question-answering
language:
- en
tags:
- spatial-reasoning
- cross-viewpoint localization
pretty_name: ViewSpatial-Bench
size_categories:
- 1K<n<10K
configs:
- config_name: ViewSpatial-Bench
  data_files:
  - split: test
    path: ViewSpatial-Bench.json
---
# **ViewSpatial-Bench: Evaluating Multi-perspective Spatial Localization in Vision-Language Models**


## Dataset Description

<!-- Provide a longer summary of what this dataset is. -->
We introduce **ViewSpatial-Bench**, a comprehensive benchmark with over 5,700 question-answer pairs across 1,000+ 3D scenes from ScanNet and MS-COCO validation sets. This benchmark evaluates VLMs' spatial localization capabilities from multiple perspectives, specifically testing both egocentric (camera) and allocentric (human subject) viewpoints across five distinct task types.

ViewSpatial-Bench addresses a critical gap: while VLMs excel at spatial reasoning from their own perspective, they struggle with perspective-taking—adopting another entity's spatial frame of reference—which is essential for embodied interaction and multi-agent collaboration. The figure below shows the construction pipeline and example demonstrations of our benchmark.

<img alt="ViewSpatial-Bench construction pipeline and example questions" src="https://cdn.jsdelivr.net/gh/lidingm/blog_img/img/202505222134833.png" style="width: 100%; max-width: 1000px;" />

The dataset contains the following fields:
| Field Name | Description |
| :--------- | :---------- |
| `question_type` | Type of spatial reasoning task, includes 5 distinct categories for evaluating different spatial capabilities |
| `image_path` | Path to the source image, includes data from two sources: `scannetv2_val` (ScanNet validation set) and `val2017` (MS-COCO validation set) |
| `question` | The spatial reasoning question posed to the model |
| `answer` | The correct answer to the question |
| `choices` | Multiple choice options available for the question |
- **Language(s) (NLP):** en
- **License:** apache-2.0

## Uses

**With HuggingFace datasets library.**
```py
from datasets import load_dataset
ds = load_dataset("lidingm/ViewSpatial-Bench")
```

## Benchmark

We provide benchmark results for various models on our benchmark. *More model evaluations will be added.*

<table>
  <thead>
    <tr>
      <th rowspan="2">Model</th>
      <th colspan="3">Camera-based Tasks</th>
      <th colspan="4">Person-based Tasks</th>
      <th rowspan="2">Overall</th>
    </tr>
    <tr>
      <th>Rel. Dir.</th>
      <th>Obj. Ori.</th>
      <th>Avg.</th>
      <th>Obj. Ori.</th>
      <th>Rel. Dir.</th>
      <th>Sce. Sim.</th>
      <th>Avg.</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td colspan="9"><em>Proprietary Models</em></td>
    </tr>
    <tr>
      <td>GPT-4o</td>
      <td>41.46</td><td>19.58</td><td>33.57</td>
      <td>42.97</td><td>40.86</td><td>26.79</td><td>36.29</td><td>34.98</td>
    </tr>
    <tr>
      <td>Gemini-2.0-Flash</td>
      <td>45.29</td><td>12.95</td><td>33.66</td>
      <td>41.16</td><td>32.78</td><td>21.90</td><td>31.53</td><td>32.56</td>
    </tr>
    <tr>
      <td>GPT-5-mini</td>
      <td>56.97</td><td>27.41</td><td>46.34</td>
      <td>43.98</td><td>49.29</td><td>26.06</td><td>38.77</td><td>42.44</td>
    </tr>
    <tr>
      <td>Gemini-2.5-Flash</td>
      <td>52.62</td><td>23.09</td><td>42.00</td>
      <td>42.97</td><td>42.16</td><td>20.27</td><td>34.22</td><td>37.99</td>
    </tr>
    <tr>
      <td>Gemini-2.5-Pro</td>
      <td>58.71</td><td>32.73</td><td>49.37</td>
      <td><u>48.59</u></td><td>45.84</td><td>25.79</td><td>39.24</td><td>44.15</td>
    </tr>
    <tr>
      <td>Gemini-3.0-Flash</td>
      <td><u>62.94</u></td><td>35.54</td><td>53.08</td>
      <td>44.88</td><td>60.69</td><td>26.24</td><td>42.40</td><td>47.58</td>
    </tr>
    <tr>
      <td>GLM-4.6v</td>
      <td>56.35</td><td>36.35</td><td>49.16</td>
      <td><b>48.90</b></td><td>47.39</td><td>23.44</td><td>38.91</td><td>43.87</td>
    </tr>
    <tr>
      <td>Doubao-Seed-1.8</td>
      <td>62.10</td><td><b>45.28</b></td><td><u>56.05</u></td>
      <td>44.98</td><td><u>62.47</u></td><td><b>33.67</b></td><td>45.74</td><td><u>50.74</u></td>
    </tr>
    <tr>
      <td>Doubao-Seed-2.0</td>
      <td><b>65.60</b></td><td><u>44.78</u></td><td><b>58.11</b></td>
      <td>47.19</td><td><b>72.09</b></td><td><u>33.57</u></td><td><b>49.20</b></td><td><b>53.52</b></td>
    </tr>
    <tr>
      <td colspan="9"><em>Open-Source General Models</em></td>
    </tr>
    <tr>
      <td>InternVL2.5 (2B)</td>
      <td>38.52</td><td>22.59</td><td>32.79</td>
      <td>47.09</td><td>40.02</td><td>25.70</td><td>37.04</td><td>34.98</td>
    </tr>
    <tr>
      <td>Qwen3-VL (4B)</td>
      <td>46.98</td><td>28.01</td><td>40.16</td>
      <td>45.68</td><td>29.22</td><td>17.74</td><td>30.48</td><td>35.17</td>
    </tr>
    <tr>
      <td>Qwen2.5-VL (7B)</td>
      <td>46.64</td><td>29.72</td><td>40.56</td>
      <td>37.05</td><td>35.04</td><td>28.78</td><td>33.37</td><td>36.85</td>
    </tr>
    <tr>
      <td>LLaVA-NeXT-Video (7B)</td>
      <td>26.34</td><td>19.28</td><td>23.80</td>
      <td>44.68</td><td>38.60</td><td>29.05</td><td>37.07</td><td>30.64</td>
    </tr>
    <tr>
      <td>LLaVA-OneVision (7B)</td>
      <td>29.84</td><td>26.10</td><td>28.49</td>
      <td>22.39</td><td>31.00</td><td>26.88</td><td>26.54</td><td>27.49</td>
    </tr>
    <tr>
      <td>InternVL2.5 (8B)</td>
      <td>49.41</td><td><b>41.27</b></td><td>46.48</td>
      <td>46.79</td><td>42.04</td><td><u>32.85</u></td><td>40.20</td><td>43.24</td>
    </tr>
    <tr>
      <td>Qwen3-VL (8B)</td>
      <td>54.60</td><td>30.32</td><td>45.87</td>
      <td>45.28</td><td>35.75</td><td>26.79</td><td>35.61</td><td>40.58</td>
    </tr>
    <tr>
      <td>Llama-3.2-Vision (11B)</td>
      <td>25.27</td><td>20.98</td><td>23.73</td>
      <td><u>51.20</u></td><td>32.19</td><td>18.82</td><td>33.61</td><td>28.82</td>
    </tr>
    <tr>
      <td>InternVL3 (14B)</td>
      <td>54.65</td><td>33.63</td><td>47.09</td>
      <td>33.43</td><td>37.05</td><td>31.86</td><td>33.88</td><td>40.28</td>
    </tr>
    <tr>
      <td>Kimi-VL-Instruct (16B)</td>
      <td>26.85</td><td>22.09</td><td>25.14</td>
      <td><b>63.05</b></td><td>43.94</td><td>20.27</td><td>41.52</td><td>33.58</td>
    </tr>
    <tr>
      <td>Qwen2.5-VL (32B)</td>
      <td>39.03</td><td>29.92</td><td>35.75</td>
      <td>36.45</td><td>34.68</td><td>21.09</td><td>30.18</td><td>32.88</td>
    </tr>
    <tr>
      <td>Qwen2.5-VL (72B)</td>
      <td>50.65</td><td>26.71</td><td>42.04</td>
      <td>42.17</td><td>42.76</td><td>24.80</td><td>35.82</td><td>38.83</td>
    </tr>
    <tr>
      <td>Qwen3-VL-Thinking (235B)</td>
      <td><u>59.73</u></td><td>36.95</td><td><u>51.54</u></td>
      <td>43.67</td><td><u>48.93</u></td><td>31.67</td><td><u>40.67</u></td><td><u>45.94</u></td>
    </tr>
    <tr>
      <td>Qwen3.5-Plus (397B)</td>
      <td><b>62.21</b></td><td><u>38.65</u></td><td><b>53.74</b></td>
      <td>50.20</td><td><b>68.17</b></td><td><b>38.37</b></td><td><b>50.90</b></td><td><b>52.28</b></td>
    </tr>
    <tr>
      <td colspan="9"><em>Multi-View Spatial Fine-Tuning</em></td>
    </tr>
    <tr>
      <td>Qwen2.5-VL (3B)</td>
      <td>43.43</td><td>33.33</td><td>39.80</td>
      <td>39.16</td><td>28.62</td><td>28.51</td><td>32.14</td><td>35.85</td>
    </tr>
    <tr>
      <td>+SFT</td>
      <td><b>83.59</b></td><td><b>87.65</b></td><td><b>85.05</b></td>
      <td><b>90.16</b></td><td><b>71.14</b></td><td><b>75.75</b></td><td><b>79.31</b></td><td><b>82.09</b></td>
    </tr>
    <tr>
      <td><em>Improvement over backbone</em></td>
      <td>+40.16</td><td>+54.32</td><td>+45.25</td>
      <td>+51.00</td><td>+42.52</td><td>+47.24</td><td>+47.17</td><td>+46.24</td>
    </tr>
    <tr>
      <td>Random Baseline</td>
      <td>25.16</td><td>26.10</td><td>25.50</td>
      <td>24.60</td><td>31.12</td><td>26.33</td><td>27.12</td><td>26.33</td>
    </tr>
  </tbody>
</table>