segrails.pipeline.postprocess

Post-processing pipeline step.

Wraps _PostProcessor to convert an inference mask into structured track detections (center points, directions, IoU scores) using Hough lines and contour analysis.

Classes

PostprocessStep

Run post-processing on a segmentation mask to extract rail track detections.

Module Contents

class segrails.pipeline.postprocess.PostprocessStep(config: altametris.segrails.config.SegrailsConfig)

Run post-processing on a segmentation mask to extract rail track detections.

Wraps _PostProcessor. All detection parameters are read from SegrailsConfig.

Parameters:

config – SegrailsConfig instance — provides inference_image_size and all detection_* parameters.

Example

>>> step = PostprocessStep(config)
>>> result = step.run(image_np, mask, geotransform, filename="tile_001")
>>> print(result.n_tracks)
_config
_postprocessor
run(image: numpy.typing.NDArray[numpy.uint8], mask: numpy.typing.NDArray[numpy.uint8], geotransform: altametris.geo3d.rasters.types.GeoTransform, output_dir: pathlib.Path | None = None, filename: str = 'detection') → altametris.segrails.result.SegrailsPrediction

Extract rail detections from a segmentation mask.

Parameters:

  • image – RGB raster image as uint8 ndarray of shape (H, W, 3), already resized to inference_image_size (as returned by InferenceStep.predict()). Passed directly to avoid re-reading the raster file that was already opened during inference.

  • mask – Binary segmentation mask of shape (H, W) with dtype uint8, as returned by InferenceStep.predict().

  • geotransform – GeoTransform from RasterStep.generate(), used to map pixel coordinates to geographic coordinates.

  • output_dir – Optional directory for saving the detection JSON and debug images. Created if it does not exist.

  • filename – Base name for output files (no extension). Defaults to "detection".

Returns:

SegrailsPrediction with n_tracks, valid_tracks, crossings, total_missed_detections, total_low_detections and enriched tracks list.

_format_predictions(raw: dict) → altametris.segrails.result.SegrailsPrediction

Build SegrailsPrediction from raw _PostProcessor output in a single pass.

Enriches each track dict with low_detections, missed_detections and overall_iou. Uses detection_iou_threshold from config.

Parameters:

raw – Raw dict from _PostProcessor.run() with a "tracks" key.

Returns:

SegrailsPrediction with all counters populated.