KubriCount: Synthetic Benchmark for Multi-Grained Visual Counting

Introduced in a 2026 paper from SJTU, KubriCount is a large-scale synthetic benchmark for multi-grained visual counting. It contains 6,736 samples and reframes open-world counting as a prompt-following problem across five explicit semantic granularity levels. The dataset is hosted by Voxel51 on the Hugging Face platform.

Use Cases

• Benchmarking visual counting models based on the described multi-grained counting tasks.
• Training models for prompt-following counting across different semantic granularities.
• Researching open-world visual understanding based on the synthetic benchmark structure.

Strengths

• 6,736 samples provide a substantial base for model training and evaluation.
• The dataset is described as supporting five explicit semantic granularity levels for counting tasks.
• It is introduced as a large-scale benchmark in a 2026 academic paper, suggesting recent development.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Voxel51 (host), introduced in a paper from SJTU (Shanghai Jiao Tong University).

Collection Method

Synthetic data generation, as described in the paper 'Count Anything at Any Granularity'.

Freshness

Last updated 2026-05-26 21:24:14.