erayyapagci/yolo26n-question-segmentation-v2

YOLO26N Question Segmentation

Smaller release checkpoint that slightly outperforms the YOLO26M release on both held-out test sets.

• Task: object detection
• Classes: question
• Input size: 1280
• Base model: yolo26n

Intended use

This model is designed to detect question regions in exam booklets, worksheets, and PDF page renders. It works best on dense educational pages with clear question boundaries.

Release notes

This release checkpoint was trained for 50 epochs on the cleaned one-class question detection dataset.

Compared with the previously published HF YOLO11M baseline, this release improves held-out mAP50-95 by +0.026 on the old test set and +0.062 on the combined test set.

Test results

Old held-out test set

Model	Size	Precision	Recall	mAP50	mAP50-95
YOLO11M Question Segmentation v1	116 MB / 20.1M params	0.962	0.980	0.979	0.898
YOLO26M Question Segmentation v2	168 MB / 21.8M params	0.990	0.982	0.988	0.920
YOLO26S Question Segmentation v2	20 MB / 9.9M params	0.988	0.983	0.990	0.923
YOLO26N Question Segmentation v2	5.3 MB / 2.5M params	0.988	0.983	0.989	0.924

Combined held-out test set

Model	Size	Precision	Recall	mAP50	mAP50-95
YOLO11M Question Segmentation v1	116 MB / 20.1M params	0.962	0.973	0.984	0.900
YOLO26M Question Segmentation v2	168 MB / 21.8M params	0.987	0.978	0.992	0.957
YOLO26S Question Segmentation v2	20 MB / 9.9M params	0.988	0.991	0.993	0.963
YOLO26N Question Segmentation v2	5.3 MB / 2.5M params	0.991	0.988	0.993	0.962

Benchmark graphs

Confidence guidance

From the confidence sweep used in this release:

• Best benchmark confidence for YOLO26S and YOLO26N is conf=0.001 (highest mAP50-95 on both held-out test sets).
• For practical inference, start with conf=0.25; use conf=0.20 if you want fewer missed borderline boxes.

Qualitative examples

The newer YOLO26 checkpoints improve most clearly in two cases:

• Less false positives on pages with no questions.
• Better recovery of harder placements such as one-row questions and questions that occupy an entire column.

Example no-question page: GT + YOLO11M Question Segmentation v1 + YOLO26M Question Segmentation v2 + YOLO26S Question Segmentation v2 + YOLO26N Question Segmentation v2. The YOLO11M panel uses the earlier low-threshold prediction pass and shows two false positives; the YOLO26 checkpoints stay empty on the same page.

Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(repo_id="erayyapagci/yolo26n-question-segmentation-v2", filename="best.pt")
model = YOLO(weights)
results = model("page.png", imgsz=1280, conf=0.25)

Direct URL loading also works:

from ultralytics import YOLO

model = YOLO("https://huggingface.co/erayyapagci/yolo26n-question-segmentation-v2/resolve/main/best.pt")
results = model("page.png", imgsz=1280, conf=0.25)

Data note

Some of the training data comes from public Roboflow projects used in earlier versions of this question-segmentation pipeline:

1. PDF Soru Cikarma (tanimazsinu): Link
2. WholeQuestionDetection (Gazi University): Link
3. ExamBuddy (ExamBuddy): Link
4. Questions (Terry Li): Link
5. Question Parsing from Document (Sefa): Link
6. Question Dedector (Nur Etinkaya): Link
7. Sorukes (Sorualgilama): Link
8. Question Detection (Cognizen): Link
9. Questions2 (Fiver): Link
10. Question-New (Question): Link

License

This release is published under CC-BY-4.0.