Postprocessors

Postprocessor is a class which processes prediction and/or annotation data after model inference and before metric calculation. Unlike adapters, postprocessors do not work with raw model output, but with specific representation format. (e. g. clip boxes postprocessor expects detection annotation and detection prediction for processing).

In case when you use complicated representation located in representation container, you can add options annotation_source and prediction_source in configuration file, if you want process only specific representations, another way postprocessor will be used for all suitable representations. annotation_source and prediction_source should contain comma separated list of annotation identifiers and output layer names respectively.

Every postprocessor has parameters available for configuration. The postprocessor and its parameters are set through the configuration file. Postprocessors are provided in datasets section of configuration file to use specific postprocessor.

Supported Postprocessors

Accuracy Checker supports following set of postprocessors:

  • cast_to_int - casting detection bounding box coordinates given in floating point format to integer. Supported representations: DetectionAnotation, DetectionPrediction, TextDetectionAnnotation, TextDetectionPrediction.

    • round_policy - method for rounding: nearest, greater, lower, nearest_to_zero.

  • clip_boxes - clipping detection bounding box sizes. Supported representations: DetectionAnotation, DetectionPrediction.

    • dst_width and dst_height - destination width and height for box clipping respectively. You can also use size instead in case when destination sizes are equal. If not provided, image size will be used.

    • apply_to - option which determines target boxes for processing (annotation for ground truth boxes and prediction for detection results, all for both).

    • bboxes_normalized is flag which says that target bounding boxes are in normalized format.

  • normalize_boxes - normalizing ground truth detection bounding boxes (cast to range [0, 1]). Supported representations: DetectionAnotation, DetectionPrediction, ActionDetectionAnotation, ActionDetectionPrediction

  • correct_yolo_v2_boxes - resizing detection prediction bbox coordinates using specific for Yolo v2 approach. Supported representations: DetectionAnotation, DetectionPrediction.

    • dst_width and dst_height - destination width and height respectively. You can also use size instead in case when destination sizes are equal.

  • resize_prediction_boxes - resizing normalized detection prediction boxes according to image size. Supported representations: DetectionAnotation, DetectionPrediction.

    • rescale if this option enabled, rescaling boxes on input size will be performed before multiplying on original input size. (Optional, default False).

    • unpadding if this option enabled, recalculation of boxes normalized coordinates to remove additional padding will be performed before multiplying on original input size. (Optional, default False).

    • uncrop if this option enabled, removes max_square crop effect while translating normalized coordinates to original image size. (Optional, default False).

  • faster_rcnn_postprocessing_resize - resizing normalized detection prediction boxes according to the original image size before preprocessing steps. Supported representations: DetectionAnotation, DetectionPrediction. At the moment works in the following cases only:

    • the preprocessing steps contains only one operation changing input image size, and the operation is resize

    • the preprocessing steps contains only two operations changing input image size, and the operations are resize and then padding.

    • rescale if this option enabled, rescaling boxes on input size will be performed before applying postprocessing (Optional, default False).

  • nms - non-maximum suppression. Supported representations: DetectionAnotation, DetectionPrediction, ActionDetectionAnnotation, ActionDetectionPrediction.

    • overlap - overlap threshold for merging detections.

    • use_min_area - boolean value to determine whether to use minimum area of two bounding boxes as base area to calculate overlap.

  • class_aware_nms - class-aware non-maximum suppression, represented as nms, which applied to boxes belonging to specific class independently. The supported representations and parameters are the same as for the nms.

  • diou_nms - distance-IoU non-maximum suppression. Supported representations: DetectionAnotation, DetectionPrediction, ActionDetectionAnnotation, ActionDetectionPrediction.

    • overlap - overlap threshold for merging detections.

  • soft_nms - soft non-maximum suppression. Supported representations: DetectionAnotation, DetectionPrediction, ActionDetectionAnnotation, ActionDetectionPrediction.

    • keep_top_k - the maximal number of detections which should be kept.

    • sigma - sigma-value for updated detection score calculation.

    • min_score - break point.

  • weighted_nms - weighted non-maximum suppression. Supported representations: DetectionAnotation, DetectionPrediction.

    • overlap - overlap threshold for merging detections.

  • filter - filtering data using different parameters. Supported representations: DetectionAnotation, DetectionPrediction.

    • apply_to - determines target boxes for processing (annotation for ground truth boxes and prediction for detection results, all for both).

    • remove_filtered - removing filtered data. Annotations support ignoring filtered data without removing as default, in other cases filtered data will be removed automatically.

    • Supported parameters for filtering: labels, min_confidence, height_range, width_range, is_empty, min_visibility, aspect_ratio, area_ratio, area_range. Filtering by height_range, width_range are also available for TextDetectionAnnotation, TextDetectionPrediction, area_range - for PoseEstimationAnnotation, PoseEstimationPrediction and TextDetectionAnnotation, TextDetectionPrediction.

  • normalize_landmarks_points - normalizing ground truth landmarks points. Supported representations: FacialLandmarksAnnotation, FacialLandmarksPrediction.

    • use_annotation_rect - allows to use size of rectangle saved in annotation metadata for point scaling instead source image size.

  • encode_segmentation_mask - encoding segmentation label image as segmentation mask. Supported representations: SegmentationAnotation, SegmentationPrediction.

    • apply_to - determines target masks for processing (annotation for ground truth and prediction for detection results, all for both). Note: this postprocessing requires a specific dataset meta: segmentation_colors for annotations and prediction_to_gt_labels for predictions.

  • resize_segmentation_mask - resizing segmentation mask. Supported representations: SegmentationAnotation, SegmentationPrediction, BackgroundMattingAnnotation, BackgroundMattingPrediction, SalientRegionAnnotation, SalientRegionPrediction, AnomalySegmentationAnnotation, AnomalySegmentationPrediction.

    • dst_width and dst_height - destination width and height for resize respectively. You can also use size instead in case when destination sizes are equal. For resizing to final input size without padding, you can use to_dst_image_size flag with value True. If any of these parameters are not specified, original image size will be used as default.

    • apply_to - determines target masks for processing (annotation for ground truth and prediction for detection results, all for both).

  • extend_segmentation_mask - extending annotation segmentation mask to predicted mask size making border filled by specific value. Supported representations: SegmentationAnotation, SegmentationPrediction.

    • filling_label - value for filling border (default 255).

    • pad_type - padding space location. Supported: center, left_top, right_bottom (Default is center).

  • zoom_segmentation_mask - zooming segmentation mask. Supported representations: SegmentationAnotation, SegmentationPrediction.

    • zoom - size for zoom operation.

  • crop_segmentation_mask - cropping 2-d, 3-d annotation mask. Supported representations: BrainTumorSegmentationAnnotation, BrainTumorSegmentationPrediction, SegmentationAnnotation, SegmentationPrediction.

    • dst_width, dst_height and dst_volume are destination width, height and volume for cropped 3D-image respectively. You can also use size instead in case when destination sizes are equal for all three dimensions.

  • crop_or_pad-segmentation_mask - performs central cropping if original mask size greater then destination size and padding in case, when source size lower than destination. Padding filling value is 0, realization - right-bottom.

    • dst_width and dst_height are destination width and height for keypoints resizing respectively. You can also use size instead in case when destination sizes are equal. Supported representations: SegmentationAnotation, SegmentationPrediction.

  • crop_padded_prediction - performs cropping if original image was padded, to the padding size. Supported representations: SegmentationAnotation, SegmentationPrediction.

  • heatmap2keypoints - extract landmark keypoints from the heatmap. Supported representations: FacialLandmarksHeatMapAnnotation, FacialLandmarksHeatMapPrediction.

  • clip_segmentation_mask - clipping segmentation mask values. Supported representations: BrainTumorSegmentationAnnotation, BrainTumorSegmentationPrediction.

    • min_value - lower bound of range.

    • max_value - upper bound of range.

  • segmentation_prediction_resample - resamples output prediction in two steps: 1) resizes it to bounding box size; 2) extends to annotation size. Supported representations: BrainTumorSegmentationAnnotation, BrainTumorSegmentationPrediction. For correct bounding box size must be set via tag boxes_file in brats_numpy converter or crop_brats preprocessor.

    • make_argmax - applies argmax operation to prediction mask after the resampling (by default False). Must be specified only one option make_argmax.

  • transform_brats_prediction - transforms prediction from WT-TC-ET format to NCR/NET-ED-ET. Sequentially fills one-channel mask with specified values for elements passing the threshold (threshold is 0.5) from each prediction channel in specified order.

    • order - specifies filling order for channels

    • values - specifies values for each channel according to new order

  • remove_brats_prediction_padding - process output prediction in two steps: 1) remove padding; 2) extends to annotation size. Supported representations: BrainTumorSegmentationAnnotation, BrainTumorSegmentationPrediction.

    • make_argmax - applies argmax operation to prediction mask (by default False).

  • extract_answers_tokens - extract predicted sequence of tokens from annotation text. Supported representations: QuestionAnsweringAnnotation, QuestionAnsweringPrediction.

    • max_answer - maximum answer length (Optional, default value is 30).

    • n_best_size - total number of n-best prediction size for the answer (Optional, default value is 20).

  • bidaf_extract_answers_tokens - extract predicted sequence of tokens from annotation text. Supported representations: QuestionAnsweringBiDAFAnnotation, QuestionAnsweringPrediction.

  • translate_3d_poses - translating 3D poses. Supported representations: PoseEstimation3dAnnotation, PoseEstimation3dPrediction. Shifts 3D coordinates of each predicted poses on corresponding translation vector.

  • resize_super_resolution - resizing super resolution predicted image. Supported representations: SuperResolutionAnotation, SuperResolutionPrediction.

    • dst_width and dst_height - destination width and height for resizing respectively. You can also use size instead in case when destination sizes are equal. If any of these parameters are not specified, gt high resolution image size will be used as default. target - select target image for resize (prediction or annotation)

  • resize_style_transfer - resizing style transfer predicted image. Supported representations: StyleTransferAnotation, StyleTransferPrediction.

    • dst_width and dst_height - destination width and height for resizing respectively.

  • crop_image - cropping image. Supported representations: ImageInpaintingAnnotation, ImageInpaintingPrediction, ImageProcessingAnnotation, ImageProcessingPrediction.

  • corner_crop_image - corner crop of the image. Supported representations: ImageInpaintingAnnotation, ImageInpaintingPrediction, ImageProcessingAnnotation, ImageProcessingPrediction.

    • dst_width and dst_height are destination width and height

    • corner_type is type of the corner crop. Options are:

      • top-left

      • top-right

      • bottom-left

      • bottom-right Default choice is top-left

  • resize - resizing image or segmentation mask. Supported representations: SegmentationAnotation, SegmentationPrediction, StyleTransferAnotation, StyleTransferPrediction, SuperResolutionAnotation, SuperResolutionPrediction, ImageProcessingAnnotation, ImageProcessingPrediction, SalientRegionAnnotation, SalientRegionPrediction, BackgroundMattingAnnotation, BackgroundMattingPrediction, AnomalySegmentationAnnotation, AnomalySegmentationPrediction.

    • dst_width and dst_height - destination width and height for resize respectively. You can also use size instead in case when destination sizes are equal. If any of these parameters are not specified, image size will be used as default.

    • apply_to - determines target masks for processing (annotation for ground truth and prediction for detection results, all for both).

    • resize_realization - parameter specifies functionality of which library will be used for resize: opencv or pillow (default pillow is used). opencv library is supported for following representations: StyleTransferAnnotation, StyleTransferPrediction, SuperResolutionAnnotation, SuperResolutionPrediction, ImageProcessingAnnotation, ImageProcessingPrediction, ImageInpaintingAnnotation, ImageInpaintingPrediction.

  • rgb_to_gray - converts reference data stored in RGB format to gray scale. Supported representations: SuperResolutionAnnotation, SuperResolutionPrediction, ImageProcessingAnnotation, ImageProcessingPrediction, StyleTransferAnnotation, StyleTransferPrediction.

  • bgr_to_gray - converts reference data stored in BGR format to gray scale. Supported representations: SuperResolutionAnnotation, SuperResolutionPrediction, ImageProcessingAnnotation, ImageProcessingPrediction, StyleTransferAnnotation, StyleTransferPrediction.

  • remove_repeats - removes repeated predicted tokens. Supported representations: MachineTranslationPrediction, MachineTranslationAnnotation.

  • to_lower_case - convert tokens to lower case. Supported representations: MachineTranslationPrediction, MachineTranslationAnnotation.

  • sr_image_recovery - restores input in YCrCb format and converts prediction to BRG or RGB format, using prediction gray channel and input Cr and Cb channels. Supported representation: SuperResolutionPrediction.

    • target_color - target color space for super resolution image - bgr and rgb are supported. (Optional, default rgb).

    • size - size of model input for recovering YCrCb image.

    • dst_width and dst_height - width and height of model input respectively for recovering YCrCb image.

  • colorization_recovery - restores BGR image from Colorization models results represented as AB-channels in LAB color space. Supported representations: ImageProcessingAnnotation, ImageProcessingPrediction.

  • argmax_segmentation_mask - translates categorical annotation segmentation mask to numerical. Supported representations: SegmentationAnnotation, SegmentationPrediction.

  • shift_labels - shifts predicted detection labels. Supported representation: DetectionPrediction.

    • offset - value for shift.

  • normalize_salience_map - converts reference salience map from [0, 255] to [0, 1] range. Supported representations: SalientRegionAnnotation, SalientRegionPrediction.

  • min_max_normalization - normalize regression data into [0, 1] with given min and max values. Supported representation: FeatureRegressionAnnotation, RegressionAnnotation, RegressionPrediction

    • min - minimal value in range, optional, default 0.

    • max - maximal value in range.

    • apply_to - determines target masks for processing (annotation for ground truth and prediction for detection results, all for both).

  • pad_signal - add left padding to reference signal if it was done for input data. Supported representations: NoiseSuppressionAnnotation, NoiseSuppressionPrediction.

  • time_series_denormalize - apply denormalization to predicted and ground truth curves. Supported representations: TimeSeriesForecastingAnnotation, TimeSeriesForecastingQuantilesPrediction.

  • interpolation - Interpolates values into target range. Supported representation: ImageProcessingPrediction, ImageProcessingAnnotation.

    • mode - interpolation mode. Currently, only linear mode available. (Optional, default value is linear)

    • target_min - Minimum of target range. (Optional, default value is 0)

    • target_max - Maximum of target range. (Optional, default value is 255)

    • as_log - Operates over logarithmic scale of source values. (Optional, default value is False)

  • invert_mask - inverts segmentation or background matting mask. Supported representations: SegmentationAnnotation, SegmentationPrediction, BackgroundMattingAnnotation, BackgroundMattingPrediction.

    • apply_to - determines target masks for processing (annotation for ground truth and prediction for detection results, all for both).

  • rescale_mask - rescale mask data range to have specified min and max values. Supported representations: BackgroundMattingAnnotation, BackgroundMattingPrediction.

    • apply_to - determines target masks for processing (annotation for ground truth and prediction for detection results, all for both).

    • min - minimal value in range, optional, default 0.

    • max - maximal value in range.

  • sentence_similarity_pooling - pool words embeddings for sentence similarity task. Supported representations: SentenceSimilarityAnnotation, ReidentificationPrediction.

    • pooling type - pooling type for embeddings - mean for mean pooling, max for max pooling (Optional, default mean).

    • remove_padding - remove end of string padding from word embeddings (Optional, default True).

  • hand_landmarks - converts hand landmark coordinates to source image coordinate space. Supported representations: HandLandmarksPrediction.