File size: 4,634 Bytes
bf053be
 
 
 
 
 
 
 
 
8a93328
 
 
b5bd302
 
 
 
 
 
 
 
 
 
 
 
4030501
 
b5bd302
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a740843
 
 
b5bd302
 
 
 
 
8a93328
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7b3fe6e
8a93328
7b3fe6e
8a93328
7b3fe6e
8a93328
7b3fe6e
8a93328
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ceebb6b
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
---
license: cc-by-4.0
task_categories:
- visual-question-answering
language:
- en
pretty_name: HandVQA
size_categories:
- 1M<n<10M
---


<div align="center">

# HandVQA: Diagnosing and Improving Fine-Grained Spatial Reasoning about Hands in Vision-Language Models

<p align="center">
  <strong>CVPR 2026</strong>
</p>

<p align="center">
  MD Khalequzzaman Chowdhury Sayem<sup>1*</sup>,
  Mubarrat Tajoar Chowdhury<sup>1*</sup>,
  Yihalem Yimolal Tiruneh<sup>1</sup>,
  Muneeb A. Khan<sup>1</sup>,
  Muhammad Salman Ali<sup>1</sup>,
  Binod Bhattarai<sup>2,3,4&dagger;</sup>,
  Seungryul Baek<sup>1&dagger;</sup>
</p>

<p align="center">
  <sup>1</sup>UNIST,
  <sup>2</sup>University of Aberdeen,
  <sup>3</sup>University College London,
  <sup>4</sup>Fogsphere (Redev.AI Ltd)
</p>

<p align="center">
  <sup>*</sup>Equal contribution.
  <sup>&dagger;</sup>These authors jointly supervised this work.
</p>

<p align="center">
  <a href="https://kcsayem.github.io/handvqa/"><img alt="Project Page" src="https://img.shields.io/badge/Project-Page-3b3b3b?style=for-the-badge"></a>
  <a href="https://arxiv.org/abs/2603.26362">
  <img alt="arXiv" src="https://img.shields.io/badge/arXiv-2603.26362-b31b1b?style=for-the-badge">
</a>
  <a href="https://huggingface.co/datasets/kcsayem/handvqa"><img alt="Dataset" src="https://img.shields.io/badge/HuggingFace-Dataset-fcc624?style=for-the-badge"></a>
  <a href="https://github.com/kcsayem/handvqa"><img alt="GitHub Code" src="https://img.shields.io/badge/GitHub-Code-2f3542?style=for-the-badge"></a>
</p>

</div>

HandVQA is a large-scale diagnostic benchmark designed to evaluate fine-grained spatial reasoning in vision-language models (VLMs), focusing on articulated hand pose understanding.

It contains over **1.6 million multiple-choice questions (MCQs)** derived from 3D hand annotations, probing joint-level relationships such as angles, distances, and relative positions.

## Dataset Description

### Motivation
Despite strong performance on general VQA tasks, VLMs struggle with fine-grained spatial reasoning, especially for articulated structures like human hands.

HandVQA is designed to diagnose these limitations by evaluating:
- Joint angle understanding
- Inter-joint distances
- Relative spatial positions (X, Y, Z)

### Data Sources
The dataset is built from:
- FreiHAND
- InterHand2.6M
- FPHA

using their 3D hand joint annotations.

## Task Format

Each sample consists of:
- An image of a hand
- A multiple-choice question (MCQ)
- 4 candidate answers
- 1 correct answer

### Subtasks
HandVQA includes 5 categories:
1. Angle
2. Distance
3. Relative Position (X-axis)
4. Relative Position (Y-axis)
5. Relative Position (Z-axis)

Each question probes a specific geometric relation between hand joints.

## Example

**Question:**
From the options below, choose the correct description.

**Options:**

A. The middle finger is bent completely inward at the distal interphalangeal joint.

B. The middle finger is bent inward at the distal interphalangeal joint.  

C. The middle finger is bent slightly inward at the distal interphalangeal joint.

D. The middle finger is straight at the distal interphalangeal joint.

**Answer:** D

## Dataset Statistics

- Total questions: ~1.6M+
- Number of datasets used: 3
- Categories: 5

## Data Generation Pipeline

HandVQA is generated using a deterministic pipeline:

1. **Pose Descriptor Extraction**
   - Compute angles, distances, and relative positions from 3D joints

2. **Discretization**
   - Convert continuous values into categories (e.g., bent, straight)

3. **Sentence Generation**
   - Fill structured templates

4. **MCQ Formation**
   - Generate correct + distractor answers
  

## Intended Uses

- Benchmarking spatial reasoning in VLMs
- Training spatially-aware multimodal models
- Evaluating hallucination in pose understanding
- Studying geometry-grounded reasoning

## Evaluation Metrics

- Accuracy
- Mean Absolute Error (MAE) for ordinal tasks (angle, distance)

HandVQA evaluates whether models truly understand spatial geometry rather than relying on language priors.

## Citation
```bibtex
@inproceedings{sayem2026handvqa,
  title     = {HandVQA: Diagnosing and Improving Fine-Grained Spatial Reasoning about Hands in Vision-Language Models},
  author    = {Sayem, MD Khalequzzaman Chowdhury and Chowdhury, Mubarrat Tajoar and Tiruneh, Yihalem Yimolal and Khan, Muneeb A. and Ali, Muhammad Salman and Bhattarai, Binod and Baek, Seungryul},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
  note      = {arXiv:2603.26362}
}
```