NYUv2: Indoor Scene Images with Depth, Segmentation, and Surface Normals

NYUv2 provides a multimodal dataset for indoor scene understanding tasks. Each sample includes an 'image', 'segmentation', 'depth', 'normal', and synthetically generated 'noise'. The dataset was uploaded by tanganke to Hugging Face on June 6, 2024, and is split into 'train' and 'val' sets.

Use Cases

• Train semantic segmentation models based on the 'segmentation' labels.
• Develop monocular depth estimation systems based on the 'depth' data.
• Benchmark surface normal prediction algorithms based on the 'normal' information.
• Test model robustness to synthetic noise based on the generated 'noise' field.
• Explore multi-task learning for scene understanding using the combined modalities.

Strengths

• Each sample provides five aligned modalities ('image', 'segmentation', 'depth', 'normal', 'noise') for multi-task learning.
• The dataset is structured with predefined 'train' and 'val' splits, facilitating standard evaluation.
• It is hosted on Hugging Face Datasets, which provides a standardized API for loading and processing.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The 'noise' is synthetically generated using torch.rand(), which may not reflect real-world noise patterns.

Provenance

Source

Original data sourced from Tsinghua Cloud, transformed by tanganke.

Collection Method

Transformed into Huggingface Dataset format.

Freshness

Last updated 2024-06-06 04:14:08; freshness should be verified.