Published October 24, 2026
| Version v1
Video/Audio
Open
VieSign-10K: A Vietnamese Isolated Sign Language Dataset with Original and Signer-Independent Benchmark Splits
Description
### VieSign-10K
The complete corpus contains 10,515 videos covering 3,659 sign classes. It provides the source collection from which the evaluation benchmarks are derived.
### VieSign-1182
The original benchmark contains 7,664 unique videos covering 1,182 classes. It uses a class-stratified split that permits signer overlap between partitions.
| Partition | Videos |
|---|---:|
| Train | 4,154 |
| Validation | 1,755 |
| Test | 1,755 |
| **Total** | **7,664** |
This protocol evaluates recognition when signer identities may be shared across training, validation, and testing.
### VieSign-SI447
The signer-independent benchmark contains 4,252 videos covering 447 classes. Signer groups are mutually disjoint, and all 447 classes appear in every partition.
| Partition | Videos | Signers | Classes | Video proportion |
|---|---:|---:|---:|---:|
| Train | 2,344 | 25 | 447 | 55.13% |
| Validation | 1,021 | 13 | 447 | 24.01% |
| Test | 887 | 12 | 447 | 20.86% |
| **Total** | **4,252** | **50** | **447** | **100.00%** |
VieSign-SI447 is derived from VieSign-1182 by retaining classes represented in all three signer-disjoint partitions. This filtering excludes 735 classes comprising 3,412 videos.
Use only this 447-class configuration for the revised benchmark. Do not combine it with the alternative 572-class configuration.