Views
No views yet
| ID | Class |
|---|---|
| 0 | frame |
| 1 | dialogue_text |
| 2 | balloon |
| 3 | onomatopoeia_text |
ShadowB/Manga109-panel-balloon-text-yolov26-segmentation
using MangaSeg annotations and locally licensed Manga109 page images. The
published repository contains no dataset images or annotations.model.safetensors: inference state dictionary; this is the only weight file.yolo26s-seg.yaml: YOLO26s segmentation architecture.config.json: class names and recommended image size.export-manifest.json: weight checksum and conversion validation.max_abs_diff = 0).1import json
2from pathlib import Path
3
4from huggingface_hub import snapshot_download
5from safetensors.torch import load_file
6from ultralytics import YOLO
7
8root = Path(
9 snapshot_download(
10 "mayocream/koharu-yolo26s",
11 allow_patterns=["*.safetensors", "*.yaml", "*.json"],
12 )
13)
14config = json.loads((root / "config.json").read_text())
15model = YOLO(root / "yolo26s-seg.yaml", task="segment")
16model.model.load_state_dict(load_file(root / "model.safetensors"), strict=True)
17model.model.names = {int(key): value for key, value in config["names"].items()}
18
19results = model.predict("page.jpg", imgsz=1280, conf=0.25).pt).| Evaluation | Box mAP50 | Box mAP50-95 | Mask mAP50 | Mask mAP50-95 |
|---|---|---|---|---|
| Four classes | 0.879 | 0.795 | 0.830 | 0.422 |
| Shared three classes | 0.975 | 0.923 | 0.946 | 0.492 |
| Upstream checkpoint, shared classes | 0.932 | 0.877 | 0.741 | 0.380 |