wuhuohua/UI_lens

Dataset Card for UI-Lens

Dataset Details

UI-Lens is a benchmark designed for evaluating UI display defect detection in commercial mobile applications.

The Chinese version of the dataset comprises 4,566 meticulously annotated UI pages from 11 leading Chinese commercial applications. Unlike existing UI benchmarks focused on clean interfaces, UI-Lens specifically targets challenging real-world scenarios including display rendering issues, boundary understanding problems, and semantic inconsistencies.

Upcoming Release: We will also be releasing an English version of the dataset containing approximately 3,000 annotated UI pages. This data is currently undergoing final cleaning and formatting and will be published soon.

• Language(s): Chinese (UI content, annotations), English (annotations)
• License: cc-by-4.0
• Paper: [To be released upon publication]

Privacy Statement

All interfaces containing sensitive personal information have been strictly removed or blurred during the data processing phase. The public release version does not contain any privacy-sensitive data. However, if you discover any inadvertently included sensitive information within the dataset, please contact us at {wxiang,yexinruiwu}@zju.edu.cn, and we will take immediate action to address and remove it.

Uses

Direct Use

UI-Lens is designed for:

• Evaluating defect detection methods - It assesses the current performance limitations of models and engineering approaches.

Out-of-Scope Use

• Training general-purpose UI understanding models - The dataset is deliberately focused on defect detection, not comprehensive UI understanding

Dataset Structure

Data Organization

The dataset is organized into two main evaluation scenarios:

Single-Interface Defect Detection (5 tasks): Text Overflow\Cropped Content\Container Overlap\Undisplayed Content\Abnormal Text Ellipsis

Data file location: The annotation file is in 'label_for_single_UIs', and the image file is in' single_UIs'.

Each single-interface sample includes:

• id: identifier
• infos.image_path: Relative path to the image file
• infos.app_name: Application name
• infos.target_problem: Defect category
• infos.label_list: Labels corresponding to each bounding box
• infos.box_list: Bounding boxes as [x, y, w, h] in pixel coordinates
• infos.image_size: Original image size [width, height]

Sequence-Interface Defect Detection : Text Inconsistency

Sequence data layout (example):

---file
seq_UIs/
└── douyin/
    └── GroupBuying-Channel_01/
        ├── imgs/
        │   ├── 01.png
        │   ├── 02.png
        │   ├── 03.png
        │   ├── 04.png
        │   ├── 05.png
        │   └── 06.png
        ├── answers.json
        └── history.json

Pattern: {APP}/{sequence_name}/imgs + answers.json + history.json.

Each sequence-interface sample includes:

• answers.json: Array of groundtruth detections with fields, pages, details, type
• history.json: Array of interaction steps with fields image, history
• imgs: Ordered page images (e.g., 01.png…), referenced by pages/image

Citation

TBD