ISU-Test/isu-challenge-dataset

Dataset Card for ISU Challenge Dataset

Dataset Summary

ISU Challenge Dataset is a multi-modal in-cabin automotive dataset with a controlled synthetic core and paired real-reference scenes.

The synthetic core contains 1,000 synchronized Blender-rendered samples with:

• RGB render
• depth (EXR and PNG)
• instance segmentation
• canny edge map
• structured scenario labels

An additional 60 paired real-reference scenes occupy sample_00000 through sample_00059. Each paired scene contains real/<stem>_real.jpg, its matching images/<stem>_sim.png, structured labels, and depth/Canny/instance-segmentation auxiliaries rendered from those labels. The original synthetic records sample_00000 through sample_00059 are relocated to sample_01000 through sample_01059 so the paired scenes lead the dataset without changing the total record count.

Each sample is linked through a scene_manifest.json entry and shares its canonical name across modalities.

Supported Tasks

This dataset can support:

• Semantic segmentation using instance segmentation masks.
• Multi-label classification for cabin state/safety/context variables from labels JSON.
• Depth estimation benchmarking with paired RGB and depth.
• Edge-focused pretraining/evaluation using canny maps.

Languages

No natural-language corpus is included. Label keys and categorical values are English identifiers.

Dataset Structure

Top-level layout:

• canny/
• depth/exr/
• depth/png/
• images/
• instance_seg/
• labels/
• real/
• scene_manifest.json

All logical records use the sample_XXXXX format. For paired records sample_00000 through sample_00059, the real photo is real/<stem>_real.jpg and its matching synthetic image is images/<stem>_sim.png. The remaining synthetic-only records are sample_00060 through sample_01059.

Manifest records retain paths and structured parameters for traceability. Per-sample annotations are provided under labels/.

Data Instances

A typical manifest record:

{
    "sample_index": 0,
    "sample_name": "sample_00000",
    "json_path": "labels/sample_00000.json",
    "image_path": "images/sample_00000_sim.png",
    "real_image_path": "real/sample_00000_real.jpg",
    "depth_path": "depth/png/sample_00000_depth.png",
    "instance_seg_path": "instance_seg/sample_00000_instance_seg.png",
    "canny_path": "canny/sample_00000_canny.png",
    "params": {
        "env": "HIGHWAY",
        "driver_gender": "FEMALE",
        "driver_safety_belt": "NO",
        "passenger_codriver": "NO",
        "suitcase": "YES",
        "driver_height_m": 1.96,
        "driver_weight_kg": 83.4,
        "light_front": 0.31,
        "env_strength": 0.92
    }
}

For a paired reference scene, the record additionally contains:

{
    "sample_name": "sample_00000",
    "real_image_path": "real/sample_00000_real.jpg"
}

Data Fields

• sample_index (int): Unique sample id.
• sample_name (string): Canonical ID, either sample_XXXXX or <timestamp>_real.
• json_path (string): Path to scene label.
• image_path (string): Path to RGB rendered image.
• real_image_path (string, optional): Path to the paired real-reference JPG.
• depth_path (string): Path to normalized PNG depth map; the matching EXR is stored under depth/exr/.
• instance_seg_path (string): Path to instance segmentation mask.
• canny_path (string): Path to canny edge image.
• params (object): Structured labels/conditions for the scene, including environment, occupants, safety-belt states, props, and continuous lighting/photometric parameters.

Data Splits

This release contains a single split (train-like full set).

Split	Samples
synthetic core	1000
paired real-reference scenes	60
all logical scene records	1060

Dataset Creation

Curation Rationale

The dataset was created to support repeatable, controlled in-cabin perception experiments where scene factors can be explicitly configured and audited.

Source Data

Initial Data Collection and Normalization

The synthetic core and paired auxiliary modalities are generated from structured scenario configurations. Paired real-reference JPGs are separately collected reference images that use the matching scenario labels. The Blender renders use:

• Renderer: Cycles
• Resolution: 960x540
• Seed: 42

Annotations

Annotation process

Synthetic annotations are generated by the simulation pipeline. Labels for paired real-reference scenes are stored as matching JSON files and mirrored in their manifest parameters.

Personal and Sensitive Information

The synthetic core contains no real-person photos. The paired real-reference subset contains real images; confirm appropriate consent, privacy review, and redistribution permissions before publishing or sharing that subset.

Considerations for Using the Data

Social Impact of Dataset

• Faster iteration for in-cabin vision models in safety-related scenarios.
• Reproducible controlled testing of feature interactions.

Discussion of Biases

Bias may arise from design choices in:

• Scenario parameter ranges.
• Asset library and occupancy patterns.
• Lighting/environment priors.

Additional Information

Dataset Curators

Lev Sorokin, Rifaath Ameen, Stefano Carlo Lambertenghi, Chen Yang.

Licensing Information

MIT License.

Citation Information

If you use this dataset, please cite:

@inproceedings{sorokin2026isutest,
  author    = {Lev Sorokin and Chen Yang and Ken E. Friedl and Andrea Stocco},
  title     = {Search-based Testing of Vision Language Models for In-Car Scene Understanding},
  booktitle = {Proceedings of the 41st IEEE/ACM International Conference on Automated Software Engineering (ASE 2026), Industry Track},
  year      = {2026},
  doi       = {10.1145/3832783.3834506}
}