Перейти к содержанию

ImageLocalizer

ImageLocalizer is a pipeline action for object detection and region localization. Training uses torchvision Faster R-CNN under the hood.

Parameters:

  • num_classes — total classes including background; when None, uses len(dataset.classes) + 1

Typical use: locate a region of interest before classification with ImageClassifier(source_mode="roi").

from hyppopipe.data.dataset import YAMLDataset
from hyppopipe.pipeline import Pipeline, Step
from hyppopipe.pipeline.image.classification import ImageClassifier
from hyppopipe.pipeline.image.localization import ImageLocalizer

pipeline = Pipeline(
    {
        "tumor_localize": Step(
            ImageLocalizer(),
            description="Locate tumor bounding box",
        ),
        "tumor_classify": Step(
            ImageClassifier(source_mode="roi"),
            inputs=["tumor_localize"],
            description="Classify tumor type inside ROI",
        ),
    },
    shift_result=False,
)

At inference time, LocalizationPrediction exposes bounding boxes and optional cropped Image views.

Documentation

Pipeline step for localization and detection (torchvision Faster R-CNN training).

__call__()

Marker callable; training uses attributes, not runtime invocation.

__init__(*, num_classes=None)

Store detection training options.

Parameters:

Name Type Description Default
num_classes int | None

Total classes including background for torchvision Faster R-CNN. When None, uses len(dataset.classes) + 1.

 None