otx.api.utils.tiler#

Tiling Module.

Classes

Tiler(tile_size, overlap, max_number, detector)

Tile Image into (non)overlapping Patches.
class otx.api.utils.tiler.Tiler(tile_size: int, overlap: float, max_number: int, detector: Model, classifier: ImageModel | None = None, segm: bool = False, mode: str = 'async', num_classes: int = 0)[source]#

Bases: object

Tile Image into (non)overlapping Patches. Images are tiled in order to efficiently process large images.

Parameters:

  • tile_size – Tile dimension for each patch

  • overlap – Overlap between adjacent tile

  • max_number – max number of prediction per image

  • detector – OpenVINO adaptor model

  • classifier – Tile classifier OpenVINO adaptor model

  • segm – enable instance segmentation mask output

  • mode – async or sync mode

crop_tile(image: ndarray, coord: List[int]) → ndarray[source]#

Crop tile from full image.

Parameters:

  • image (np.ndarray) – full-res image

  • coord (List) – tile coordinates

Returns:

cropped tile

Return type:

np.ndarray

static detection2tuple(detections: ndarray)[source]#

Convert detection to tuple.

Parameters:

detections (np.ndarray) – prediction results in numpy array

Returns:

scores between 0-1 labels (np.ndarray): label indices boxes (np.ndarray): boxes

Return type:

scores (np.ndarray)

filter_tiles_by_objectness(image: ndarray, tile_coords: List[List[int]], confidence_threshold: float = 0.35)[source]#

Filter tiles by objectness score by running tile classifier.

Parameters:

  • image (np.ndarray) – full size image

  • tile_coords (List[List[int]]) – tile coordinates

Returns:

tile coordinates to keep

Return type:

keep_coords

get_tiling_saliency_map_from_segm_masks(detections: Tuple | ndarray) → List[source]#

Post process function for saliency map of OTX MaskRCNN model for tiling.

merge_features(features: List, predictions: Tuple | ndarray) → Tuple[None, None] | List[ndarray][source]#

Merge tile-level feature vectors to image-level features.

Parameters:

  • features – tile-level features.

  • predictions – predictions with masks for whole image.

Returns:

Merged feature vector for entire image. image_saliency_map (List): Merged saliency map for entire image

Return type:

image_vector (np.ndarray)

merge_maps(features: List) → ndarray[source]#

Merge tile-level saliency maps to image-level saliency map.

Parameters:

  • features – tile-level features ((vector, map: np.array), tile_meta).

  • detected. (Each saliency map is a list of maps for each detected class or None if class wasn't) –

Returns:

Merged saliency maps for entire image.

Return type:

merged_maps (np.ndarray)

merge_results(results: List[Dict], shape: List[int])[source]#

Merge results from tiles.

Parameters:

  • results (List[Dict]) – list of tile results

  • shape (List[int]) – original full-res image shape

merge_vectors(features: List) → ndarray[source]#

Merge tile-level feature vectors to image-level feature vector.

Parameters:

features – tile-level features.

Returns:

Merged vectors for entire image.

Return type:

merged_vectors (np.ndarray)

postprocess_tile(predictions: DetectionResult, offset_x: int, offset_y: int) → Dict[str, List][source]#

Postprocess single tile prediction.

Parameters:

  • predictions (Union[List, Tuple]) – predictions from model

  • offset_x (int) – tile offset in x direction

  • offset_y (int) – tile offset in y direction

Returns:

postprocessed predictions - bboxes and masks

Return type:

Dict[str, List]

predict(image: ndarray, mode: str = 'async')[source]#

Predict by cropping full image to tiles.

Parameters:

image (np.ndarray) – full size image

Returns:

prediction results features: saliency map and feature vector

Return type:

detection

predict_async(image: ndarray, tile_coords: List[List[int]])[source]#

Predict by cropping full image to tiles asynchronously.

Parameters:

  • image (np.ndarray) – full size image

  • tile_coords (List[List[int]]) – tile coordinates

Returns:

prediction results features: saliency map and feature vector

Return type:

detection

predict_sync(image: ndarray, tile_coords: List[List[int]])[source]#

Predict by cropping full image to tiles synchronously.

Parameters:

  • image (np.ndarray) – full size image

  • tile_coords (List[List[int]]) – tile coordinates

Returns:

prediction results features: saliency map and feature vector

Return type:

detection

resize_masks(masks: List, dets: ndarray, shape: List[int])[source]#

Resize Masks.

Parameters:

  • masks (List) – list of raw np.ndarray masks

  • dets (np.ndarray) – detections including labels, scores, and boxes

  • shape (List[int]) – original full-res image shape

tile(image: ndarray) → List[List[int]][source]#

Tiles an input image to either overlapping, non-overlapping or random patches.

Parameters:

image – Input image to tile.

Returns:

Tiles coordinates