Preprocessors

Preprocessor is a class which processes input data before model inference. Every preprocessor has parameters available for configuration. The preprocessor and its parameters are set through the configuration file. Preprocessors are provided in datasets section of configuration file to use specific preprocessor.

Supported Preprocessors

Accuracy Checker supports following set of preprocessors:

• resize - resizing the image to a new width and height.

  • dst_width and dst_height are destination width and height for image resizing respectively. You can also use size instead in case when destination sizes are equal for both dimensions.

  • resize_realization - parameter specifies functionality of which library will be used for resize: opencv, pillow or tf (default opencv is used). For enabling tf you need to install TensorFlow first. For compatibility with previous releases you can also use boolean constants for selection resizing backend:

    • use_pillow parameter specifies usage of Pillow library for resizing.

    • use_tensorflow parameter specifies usage of TensorFlow Image for resizing. Requires TensorFlow installation. Accuracy Checker uses OpenCV as default image reader.

  • interpolation specifies method that will be used. Possible values depend on image processing library:

    • OpenCV : Nearest, Linear, Cubic, Area, Max, Lanczos4, Bits, Bits32

    • Pillow : None, Nearest, Cubic, Bicubic, Box, Bilinear, Lanczos, Antialias, Hamming, Linear

    • TensorFlow : Bilinear, Area, Bicubic Linear used as default for OpenCV, Bilinear as default for Pillow and TensorFlow.

  • aspect_ratio_scale allows resize with changing or saving image aspect ratio. May be done using one of these ways:

    • width - rescale width (height has fixed size, provided as dst_height or size, width size will be rescaled to save aspect ratio).

    • height - rescale height (width has fixed size, provided as dst_width or size, height size will be rescales to save aspect ratio).

    • greater - rescale greater from image sizes (smaller dimension has fixed size, greater will be rescaled to save aspect ratio)

    • fit_to_window - adaptive resize keeping aspect ratio for fit image into window with fixed size [dst_height x dst_width], but trying to make the image as big as possible.

    • frcnn_keep_aspect_ratio - adaptive resize keeping aspect ratio for fit image into window with fixed size [max_size x max_size], but trying to make the minimal dimension of image to be equal to min_size or as close to min_size as possible, where min_size = min(dst_width, dst_height), max_size = max(dst_width, dst_height).

    • ctpn_keep_aspect_ratio - adaptive resize keeping aspect ratio for fit image into window with fixed size [max_size x min_size] using this algorithm:

      1. Try to resize min original image size to minimal destination size.

      2. If scaled max size greater than maximal destination size, rescale minimal size to get max size equal to max destination size.

    • east_keep_aspect_ratio - adaptive resize keeping aspect ratio using this algorithm:

      1. If max image size greater max destination size, make max image size equal to max destination size.

      2. Make image height and width divisible by min destination size without remainder.

    • min_ratio - rescale width and height according to minimal ratio source_size / destination_size.

    • mask_rcnn_benchmark_aspect_ratio - rescale image size according preprocessing for maskrcnn-benchmark models in ONNX zoo

    • ppcrnn_ratio - calculate scales in the following way:

      1. find original image ratio (input_width / input_height)

      2. if dst_height * ratio larger then dst_width, then dst_width = 32 * ratio

      3. Otherwise dst_width = dst_height * ratio

    • ppocr_max_ratio - calculate scales in the following way:

      1. Maximum between provided destination size used as max_limit for image ddimensions.

      2. If maximum image size before resize greater max limit, ratio calculated as max_limit / max_size, otherwise ratio is equal 1.

    • ppocr_cls_ratio - calculate scales in following way:

      1. ratio calculated as image_width / image_height, destination size for height used as constant

      2. if ratio multiplied on dst_height greater then dst_width, then dst_width used for width resize, otherwise resized width is dst_height * ratio.

• factor - destination size for aspect ratio resize must be divisible by a given number without remainder. Please pay attention that this parameter only works with aspect_ratio_scale parameters.

• auto_resize - automatic resizing image to input layer shape. (supported only for one input layer case, use OpenCV for image resize)

• normalization - changing the range of pixel intensity values.

  • mean values which will be subtracted from image channels. You can specify one value for all channels or list of comma separated channel-wise values.

  • std specifies values, on which pixels will be divided. You can specify one value for all channels or list of comma separated channel-wise values. These parameters support work with precomputed values of frequently used datasets (e.g. cifar10 or imagenet).

  • images_only - prevent usage normalization for non-image inputs in multi input mode (Optional, default False).

• resize3d - resizing 3d image (e.g. MRI scans) to new size:

  • size in format (H,W,D). All values will be interpolated with 1st-order spline.

• crop_brats - performing crop of 3d images (e.g. MRI scans) by cropping all non-zero voxels. Also sets bounding boxes for segmentation_prediction_resample preprocessor (see Postprocessors)

• normalize_brats - normalization of 3d images (e.g. MRI scans) with z-score normalization

  • masked - specifies type of masking:

    • none for not applying mask

    • ignore for ignoring “empty” voxels in statistic calculation

    • nullify for nullifying initially “empty” voxels at the end

    • all for ignore and nullify

  • cutoff - cuts minimum and value to -cutoff and cutoff respectively

  • shift_value - adds to all values

  • normalize_value - divides all values

• swap_modalities - swapping modalities of MRI scan (works as channel swapping)

  • modality_order - new order

• bgr_to_rgb - reversing image channels. Convert image in BGR format to RGB.

• bgr_to_gray - converting image in BGR to gray scale color space.

• rgb_to_bgr - reversing image channels. Convert image in RGB format to BGR.

• rgb_to_gray - converting image in RGB to gray scale color space.

• bgr_to_yuv - converting image in BGR to YUV.

  • split_channels - split image channels to independent input data after conversion (Optional, default False).

  • shrink_uv - resize uv-channels in 1:2 resolution to y-channel (Optional, available only with split_channels combination).

• rgb_to_yuv - converting image in RGB to YUV.

  • split_channels - split image channels to independent input data after conversion (Optional, default False).

  • shrink_uv - resize uv-channels in 1:2 resolution to y-channel (Optional, available only with split_channels combination).

• bgr_to_nv12 - converting BGR image to NV12 format.

• rgb_to_nv12 - converting RGB image to NV12 format.

• nv12_to_bgr - converting NV12 data to BGR format.

• nv12_to_rgb - converting NV12 data to RGB format.

• bgr_to_ycrcb - converting image in BGR to YCrCb.

  • split_channels - split image channels to independent input data after conversion (Optional, default False).

• rgb_to_ycrcb - converting image in RGB to YCrCb.

  • split_channels - split image channels to independent input data after conversion (Optional, default False).

• bgr_to_lab - converts image in RGB format to LAB.

• rgb_to_lab - converts image in BGR format to LAB.

• select_channel - select channel only one specified channel from multichannel image.

  • channel - channel id in image (e.g. if you read image in RGB and want to select green channel, you need to specify 1 as channel)

• flip - image mirroring around specified axis.

  • mode specifies the axis for flipping (vertical or horizontal).

  • merge_with_original - allows addition flipped image to original (Optional, default False, original image will be replaced with flipped).

• crop - central cropping for image.

  • dst_width and dst_height are destination width and height for image resizing respectively. You can also use size instead in case when destination sizes are equal, central_fraction to define fraction of size to crop (float value (0, 1])) or max_square for cropping central part for image by minimal image size (True value for enabling this feature).

  • use_pillow parameter specifies usage of Pillow library for cropping.

• crop_rectangle - cropping region of interest using coordinates given as annotation metadata.

• extend_around_rect - scaling region of interest using annotation metadata.

  • augmentation_param is scale factor for augmentation.

• point_alignment - aligning keypoints stored in annotation metadata.

  • draw_points - allows visualize points.

  • normalize - allows to use normalization for keypoints.

  • 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.

• corner_crop - Corner crop of the image.

  • 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

• crop_or_pad - performs central cropping if original image 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.

• crop_image_with_padding - crops to center of image with padding then scales image size.

  • size - destination image height/width dimension,

  • crop_padding - the padding size to use when centering the crop.

• padding - padding for image.

  • stride - stride for padding.

  • pad_value - value for filling space around original image (should be provided as comma-separatd channel-wise float values or background for estimation background pixel color for image).

  • dst_width and dst_height are destination width and height for padded image respectively. You can also use size instead in case when destination sizes are equal for both dimensions.

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

  • use_numpy - allow to use numpy for padding instead default OpenCV.

  • numpy_pad_mode - if using numpy for padding, numpy padding mode, including constant, edge, mean, etc. (Default is constant).

  • enable_resize - allow resize image to destination size, if source image greater that destination (Optional, default False).

• tiling - image tiling.

  • margin - margin for tiled fragment of image.

  • dst_width and dst_height are destination width and height of tiled fragment respectively. You can also use size instead in case when destination sizes are equal for both dimensions.

• crop3d - central cropping for 3D data.

  • 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.

• normalize3d - normalizing 3D-images using mean and std values per channel of current image for subtraction and division respectively.

• tf_convert_image_dtype - cast image values to floating point values in range [0, 1]. Requires TensorFlow installation.

• decode_by_vocabulary - Decode words to set of indexes using model vocab.

  • vocabulary_file - path to vocabulary file for decoding. Path can be prefixed with --models argument.

  • unk_index - index of unknown symbol in vocab.

• decode_by_sentence_piece_bpe_tokenizer - Decode words to set of indexes using SentencePieceBPETokenizer.

  • vocabulary_file - path to vocabulary file for decoding. Path can be prefixed with --models argument.

  • merges_file - path to merges file for decoding. Path can be prefixed with --models argument.

  • sos_symbol - string representation of start_of_sentence symbol (default= <s>).

  • eos_symbol - string representation of end_of_sentence symbol (default= </s>).

  • add_symbols - add sos/eos symbols to sentence (default=True).

• pad_with_eos - supplement the input sequence to a specific size using a line terminator character or index.

  • eos_symbol or eos_index - line terminator symbol or index of this symbol in vocab for encoded sequence respectively.

  • sequence_len - length of sequence after supplement.

• centernet_affine_transform - CenterNet affine transformation, used for image resizing.

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

  • scale - scale factor for image (default is 1).

• free_form_mask - Applies free-form mask to the image.

  • parts - Number of parts to draw mask.

  • max_brush_width - Maximum brush width to draw mask.

  • max_length - Maximum line length to draw mask.

  • max_vertex - Maximum number vertex to draw mask.

  • inverse_mask - Allows mask inversion (1 - real image, 0 - masked area). Optional, default False (0 - real image, 1- masked area).

  • concat_mask - Allows concatenate generated mask and image to one tensor by channels. Optional, default False.

• rect_mask - Applies rectangle mask to the image.

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

  • inverse_mask - Allows mask inversion (1 - real image, 0 - masked area). Optional, default False (0 - real image, 1- masked area).

  • concat_mask - Allows concatenate generated mask and image to one tensor by channels. Optional, default False.

• custom_mask - Applies masks from custom mask dataset.

  • mask_dir - path to mask dataset to be used for inpainting.

  • inverse_mask - inverse mask before apply

  • concat_mask - Allows concatenate generated mask and image to one tensor by channels. Optional, default False.

  • mask_loader - which reader to use to load masks. The following readers can be used:

    • opencv_imread - read images using OpenCV library. Default color space is BGR.

    • pillow_imread - read images using Pillow library. Default color space is RGB.

    • scipy_imread - read images using similar approach as in scipy.misc.imread.

    • numpy_reader - read numpy dumped files.

    • tf_imread - read images using TensorFlow. Default color space is RGB. Requires TensorFlow installation.

    • inverse_mask - Allows mask inversion (1 - real image, 0 - masked area). Optional, default False (0 - real image, 1- masked areaa).

• warp_affine - warp affine transformation. (supported only with OpenCV)

  • src_landmarks - source landmarks to set as markers for the warp affine transformation.

  • dst_landmarks - destination and target landmarks to transform src_landmarks to.

  • dst_height - destination height size.

  • dst_width - destination width size.

• resample_audio - converts audio to new sample rate

  • sample_rate - sets new sample rate

• clip_audio - slices audio into several parts with equal duration

  • duration - sets duration of each clip in seconds or samples (use samples suffix), e.g. 1.5, 16000samples

  • overlap - sets overlapping for clips in percents or samples (use % or samples suffixes respectively) (no overlapping by default), e.g. 25%, 4000samples

  • max_clips - sets the maximum number of clips (clips all record by default)

• audio_normalization - normalize audio record with mean sample subtraction and division on standard deviation of samples.

  • per_frame - calculate mean and std for each frame separately (Optional, default False).

• audio_to_mel_spectrogram - performs all needed preprocessing to calculate MEL spectrogram from time-domain audio signal

  • window_size - size of time-domain signal frame, seconds

  • window_stride - intersection of frames in time-domain, seconds

  • window - weighting window type, possible choices:

    • hann - applies Hanning window to each signal frame

    • hamming - applies Hamming window to each signal frame

    • blackman - applies Blackman window to each signal frame

    • bartlett - applies Bartlett window to each signal frame

    • none - no window

  • n_fft - STFT base, samples

  • n_filt - number of MEL filters

  • splicing - number of sequentially concastenated MEL spectrums

  • sample_rate - audio sampling frequency, Hz

  • pad_to - desired length of features

  • preemph - preemph factor

  • log - applies log() to MEL features values

  • use_deterministic_dithering - Controls dithering mode:

    • True - there are no dithering in time-domain, fixed value from dither parameter added to signal spectrum

    • False - dithering in time-domain, random values with dither magnitude added to signal spectrum

  • dither - dithering value

  • stft_padded - enables padding at the end of input signal for Short Time Fourier Transform (STFT)

  • stft_boundary - specifies how to generate boundary values for Short Time Fourier Transform (STFT)

  • do_transpose - enables input transpose

• audio_patches - split audio signal on patches with specified size for multi inference processing. If input signal can not be divided by size without remainder, signal will be padded by zeros left side.

  • size - patch size.

• context_window - add context window padding to input signal.

  • cw_l - left side context window padding.

  • cw_r - right side context window padding.

  • to_multi_infer - prepare data for multi inference (Optional, default False).

• similarity_transform_box - apply to image similarity transformation to get rectangle region stored in annotation metadata/

  • box_scale - box scale factor (Optional, default 1).

  • dst_width and dst_height are destination width and height for transformed image respectively. You can also use size instead in case when destination sizes are equal for both dimensions.

• face_detection_image_pyramid - image pyramid for face detection

  • min_face_ratio - minimum face ratio to image size.

  • resize_scale - scale factor for pyramid layers.

• candidate_crop - crops candidates detected in previous stage model from input image with vertical and horizontal scaling.

  • scale_width - value to scale width relative to the original candidate width.

  • scale_height - value to scale height relative to the original candidate height.

• object_crop_with_scale - crop region from image using center coordinate and scale from annotation.

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

• one_hot_encoding - create label map based on array of indexes (analog scatter).

  • value - number for encoding label.

  • base - number for encoding other classes.

  • axis - axis responsible for classes.

  • number_of_classes - number of used classes.

• pack_raw_image - pack raw image to [H, W, 4] normalized image format with black level removal.

  • black_level - black level on the input image.

  • ratio - exposure scale ratio, optional, can be replaced by value from annotation if not provided.

  • 9-channels - for packing 9 channels images (Optional, default False).

• alpha - extracts alpha-channel data from the image.

  • channel - number of channel to extract (Optional, default 3).

• trimap - concatenates image data with alpha-channel based information for cut, keep and calculation zones in image.

  • cut_treshold - maximum level of alpha values in cut zone. Optional, default is 0.1.

  • keep_treshold - minimum level of alpha values in keep zone. Optional, default is 0.9. Pixels with alpha-channel values between cut_threshold and keep_treshold are in calculation zone.

• jpeg_compression - compress the image into jpeg format.

  • quality_factor - quality of compression, from 0 to 100 (the higher is the better).

• transpose - transpose data using specified axes order.

  • axes - list of dimensions in transposing order.

Optimized preprocessing via OpenVINO Inference Engine

OpenVINO™ is able perform preprocessing during model execution. For enabling this behaviour you can use command line parameter --ie_preprocessing True. When this option turn on, specified in config preprocessing will be translated to Inference Engine PreProcessInfo API. Note : This option is available only for dlsdk launcher and not all preprocessing operations can be ported to Inference Engine. Supported preprocessing:

• Resizing: resize without aspect_ratio_scale and with BILINEAR or AREA interpolation. Destination size is model input shape. (auto_resize is also can be used for resize with bilinear interpolation.)

• Color conversion: bgr_to_rgb, rgb_to_bgr, nv12_to_bgr, nv12_to_rgb

• Normalization: normalization with per channel mean and std