3 Postprocessor is function which processes prediction and/or annotation data after model infer and before metric calculation. For correct work postprocessors require specific representation format.
4 (e. g. clip boxes postprocessor expects detection annotation and detection prediction for processing).
6 In case when you use complicated representation located in representation container, you can add options `annotation_source` and `prediction_source` in configuration file,
7 if you want process only specific representations, another way postprocessor will be used for all suitable representations. `annotation_source` and `prediction_source` should contain
8 comma separated list of annotation identifiers and output layer names respectively.
10 Every postprocessor has parameters available for configuration.
12 Accuracy Checker supports following set of postprocessors:
14 * `cast_to_int` - casting detection bounding box coordinates given in floating point format to integer. Supported representations: `DetectionAnotation`, `DetectionPrediction`, `TextDetectionAnnotation`, `TextDetectionPrediction`.
15 * `round_policy` - method for rounding: `nearest`, `greater`, `lower`, `nearest_to_zero`.
16 * `clip_boxes` - clipping detection bounding box sizes. Supported representations: `DetectionAnotation`, `DetectionPrediction`.
17 * `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.
18 * `apply_to` - option which determines target boxes for processing (`annotation` for ground truth boxes and `prediction` for detection results, `all` for both).
19 * `bboxes_normalized` is flag which says that target bounding boxes are in normalized format.
20 * `correct_yolo_v2_boxes` - resizing detection prediction bbox coordinates using specific for Yolo v2 approach. Supported representations: `DetectionAnotation`, `DetectionPrediction`.
21 * `dst_width` and `dst_height` - destination width and height respectively. You can also use `size` instead in case when destination sizes are equal.
22 * `encode_segmentation_mask` - encoding segmentation label image as segmentation mask. Supported representations: `SegmentationAnotation`, `SegmentationPrediction`.
23 * `resize_prediction_boxes` - resizing normalized detection prediction boxes according to image size. Supported representations: `DetectionAnotation`, `DetectionPrediction`.
24 * `resize_segmentation_mask` - resizing segmentation mask. Supported representations: `SegmentationAnotation`, `SegmentationPrediction`.
25 * `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.
26 If any of these parameters are not specified, image size will be used as default.
27 * `apply_to` - determines target boxes for processing (`annotation` for ground truth boxes and `prediction` for detection results, `all` for both).
28 * `nms` - non-maximum suppression. Supported representations: `DetectionAnotation`, `DetectionPrediction`.
29 * `overlap` - overlap threshold for merging detections.
30 * `filter` - filtering data using different parameters. Supported representations: `DetectionAnotation`, `DetectionPrediction`.
31 * `apply_to` - determines target boxes for processing (`annotation` for ground truth boxes and `prediction` for detection results, `all` for both).
32 * `remove_filtered` - removing filtered data. Annotations support ignoring filtered data without removing as default, in other cases filtered data will be removed automatically.
33 * Supported parameters for filtering: `labels`, `min_confidence`, `height_range`, `width_range`, `is_empty`, `min_visibility`, `aspect_ratio`, `area_ratio`, `area_range`.
34 Filtering by `height_range`, `width_range` are also available for `TextDetectionAnnotation`, `TextDetectionPrediction`, `area_range` - for `PoseEstimationAnnotation`, `PoseEstimationPrediction` and `TextDetectionAnnotation`, `TextDetectionPrediction`.
35 * `normalize_landmarks_points` - normalizing ground truth landmarks points. Supported representations: `FacialLandmarksAnnotation`, `FacialLandmarksPrediction`.
36 * `use_annotation_rect` - allows to use size of rectangle saved in annotation metadata for point scaling instead source image size.
37 * `extend_segmentation_mask` - extending annotation segmentation mask to predicted mask size making border filled by specific value. Supported representations: `SegmentationAnotation`, `SegmentationPrediction`.
38 * `filling_label` - value for filling border (default 255).
39 * `zoom_segmentation_mask` - zooming segmentation mask. Supported representations: `SegmentationAnotation`, `SegmentationPrediction`.
40 * `zoom` - size for zoom operation.