CVPR-SignEval2026 — S23 winning ensemble (🥇 1st place)
Model checkpoints for the S23 super-ensemble that won the CVPR 2026 Isolated Sign Language Recognition (ISLR) challenge.
🏆 S23:
0.93101private /0.92165public — 1st place. 17 configs × 5 folds across 4 CNN families / 5 backbone variants (EfficientNet-B1/B2, ConvNeXt-tiny, MobileOne-S2, RegNetY-008), combined by CV-weighted probability averaging.
- Code / training / inference: https://github.com/CristianLazoQuispe/CVPR-SignEval2026
- License: MIT
What's in this repo
83 fold-level checkpoints (best.pt), organized as <config>/fold<k>/best.pt, plus each
config's config.yaml and experiment_summary.json. (17 configs × 5 folds = up to 85;
the official S23 used 83 — r15_regy_cm contributed 3/5 folds.)
| Backbone | Configs |
|---|---|
| EfficientNet-B1 | r9_both_trip · r9_cutmix_ls02 · r9_cutmix_p08 · r9_cutmix_a20 · r12_01…r12_05 · r15_b1_s43 |
| EfficientNet-B2 | r15_b2_cm_trip_unified |
| ConvNeXt-tiny | r9_cnxt_trip · r15_cnxt_cm |
| MobileOne-S2 | r13_mobileone · r14_mob_cm2_trip |
| RegNetY-008 | r13_regnety · r15_regy_cm |
Results — ensemble scaling (Kaggle private LB)
| Submission | Configs | Models | Architectures | Private LB |
|---|---|---|---|---|
| S15 | 6 | 30 | 1 | 0.92612 |
| S16 | 4 | 20 | 1 | 0.92816 |
| S17 | 9 | 45 | 1 | 0.92897 |
| S22 | 15 | 75 | 4 | 0.93019 |
| 🥇 S23 | 17 | 83 | 5 | 0.93101 |
Usage
Download all checkpoints and run inference + CV-weighted ensembling with the code repo:
from huggingface_hub import snapshot_download
local = snapshot_download(repo_id="CristianLazoQuispe/CVPR-SignEval2026-models", repo_type="model")
# -> place under experiments/ and run scripts/build_s23.sh (see code repo)
Each checkpoint is a standard PyTorch state_dict saved with torch.save. The matching
architecture/preprocessing for every config lives in its config.yaml. See the code repo's
README ("Reproduce the winning ensemble") for the end-to-end inference → ensemble pipeline.
Reproducibility
All models trained with seed: 42, deterministic: true, amp: false for bit-reproducible runs.
Citation
See the code repository for the BibTeX entry.