Refined Annotations with Soft Labels for Four Object Detection Benchmarks

Refined annotations for four popular object detection benchmark datasets: COCO 2017 validation, Pascal VOC 2012 training and validation, KITTI 2D object detection training, and Cityscapes training and validation data. Each annotation includes a bounding box, a postprocessed soft label, and original annotator responses with answer times. The dataset was published by an anonymous author via Harvard Dataverse in May 2026.

Use Cases

• Training object detection models with uncertainty-aware labels based on the provided soft label annotations.
• Analyzing annotator agreement and response times for bounding box tasks using the original annotator responses.
• Benchmarking model performance across multiple standard datasets (COCO, Pascal VOC, KITTI, Cityscapes) with refined annotations.
• Studying the impact of label refinement and soft labeling on model robustness and generalization.

Strengths

• Annotations cover four major benchmark datasets: COCO 2017, Pascal VOC 2012, KITTI, and Cityscapes.
• Includes soft label annotations ('postprocessed_softlabel') which may capture label uncertainty.
• Contains original annotator responses and answer times, allowing for analysis of annotation process.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license information are unknown, which may limit suitability assessment.
• Data may reflect the geographic, temporal, or scene biases inherent to the original source benchmark datasets.

Provenance

Source

Harvard Dataverse

Collection Method

Refined annotations derived from four popular object detection benchmark datasets.

Time Range

Temporal coverage is tied to the original benchmark datasets (e.g., COCO 2017).

Freshness

Last updated 2026 05 04 23:25:43; freshness should be verified.

Geography

Spatial coverage varies by source dataset (e.g., Cityscapes covers street scenes from multiple cities).