diff options
| -rw-r--r-- | arm_compute/core/Types.h | 128 | ||||
| -rw-r--r-- | arm_compute/runtime/CPP/CPPFunctions.h | 1 | ||||
| -rw-r--r-- | arm_compute/runtime/CPP/functions/CPPDetectionOutputLayer.h | 100 | ||||
| -rw-r--r-- | src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp | 585 | ||||
| -rw-r--r-- | utils/TypePrinter.h | 79 |
5 files changed, 893 insertions, 0 deletions
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h index 6833a66cd9..75b38c5cb8 100644 --- a/arm_compute/core/Types.h +++ b/arm_compute/core/Types.h @@ -995,6 +995,134 @@ private: std::array<float, 2> _steps; }; +/** Available Detection Output code types */ +enum class DetectionOutputLayerCodeType +{ + CORNER, /**< Use box corners */ + CENTER_SIZE, /**< Use box centers and size */ + CORNER_SIZE, /**< Use box centers and size */ + TF_CENTER /**< Use box centers and size but flip x and y co-ordinates */ +}; + +/** Detection Output layer info */ +class DetectionOutputLayerInfo final +{ +public: + /** Default Constructor */ + DetectionOutputLayerInfo() + : _num_classes(), + _share_location(), + _code_type(DetectionOutputLayerCodeType::CORNER), + _keep_top_k(), + _nms_threshold(), + _top_k(), + _background_label_id(), + _confidence_threshold(), + _variance_encoded_in_target(false), + _eta(), + _num_loc_classes() + { + _num_loc_classes = _share_location ? 1 : _num_classes; + } + /** Constructor + * + * @param[in] num_classes Number of classes to be predicted. + * @param[in] share_location If true, bounding box are shared among different classes. + * @param[in] code_type Type of coding method for bbox. + * @param[in] keep_top_k Number of total bounding boxes to be kept per image after NMS step. + * @param[in] nms_threshold Threshold to be used in NMS. + * @param[in] top_k (Optional) Number of boxes per image with top confidence scores that are fed into the NMS algorithm. Default set to -1. + * @param[in] background_label_id (Optional) Background label ID. If there is no background class, set it as -1. + * @param[in] confidence_threshold (Optional) Only consider detections whose confidences are larger than a threshold. Default set to -FLT_MAX. + * @param[in] variance_encoded_in_target (Optional) If true, variance is encoded in target. Otherwise we need to adjust the predicted offset accordingly.Default set to false. + * @param[in] eta (Optional) Eta. + */ + DetectionOutputLayerInfo(int num_classes, bool share_location, DetectionOutputLayerCodeType code_type, int keep_top_k, float nms_threshold, int top_k = -1, int background_label_id = -1, + float confidence_threshold = std::numeric_limits<float>::lowest(), bool variance_encoded_in_target = false, float eta = 1) + : _num_classes(num_classes), + _share_location(share_location), + _code_type(code_type), + _keep_top_k(keep_top_k), + _nms_threshold(nms_threshold), + _top_k(top_k), + _background_label_id(background_label_id), + _confidence_threshold(confidence_threshold), + _variance_encoded_in_target(variance_encoded_in_target), + _eta(eta), + _num_loc_classes() + { + _num_loc_classes = _share_location ? 1 : _num_classes; + } + /** Get num classes. */ + int num_classes() const + { + return _num_classes; + } + /** Get share location. */ + bool share_location() const + { + return _share_location; + } + /** Get detection output code type. */ + DetectionOutputLayerCodeType code_type() const + { + return _code_type; + } + /** Get if variance encoded in target. */ + bool variance_encoded_in_target() const + { + return _variance_encoded_in_target; + } + /** Get the number of total bounding boxes to be kept per image. */ + int keep_top_k() const + { + return _keep_top_k; + } + /** Get nms threshold. */ + float nms_threshold() const + { + return _nms_threshold; + } + /** Get eta. */ + float eta() const + { + return _eta; + } + /** Get background label ID. */ + int background_label_id() const + { + return _background_label_id; + } + /** Get confidence threshold. */ + float confidence_threshold() const + { + return _confidence_threshold; + } + /** Get top K. */ + int top_k() const + { + return _top_k; + } + /** Get number of location classes. */ + int num_loc_classes() const + { + return _num_loc_classes; + } + +private: + int _num_classes; + bool _share_location; + DetectionOutputLayerCodeType _code_type; + int _keep_top_k; + float _nms_threshold; + int _top_k; + int _background_label_id; + float _confidence_threshold; + bool _variance_encoded_in_target; + float _eta; + int _num_loc_classes; +}; + /** Pooling Layer Information class */ class PoolingLayerInfo { diff --git a/arm_compute/runtime/CPP/CPPFunctions.h b/arm_compute/runtime/CPP/CPPFunctions.h index 9c17d9e404..63df437d11 100644 --- a/arm_compute/runtime/CPP/CPPFunctions.h +++ b/arm_compute/runtime/CPP/CPPFunctions.h @@ -26,6 +26,7 @@ /* Header regrouping all the CPP functions */ #include "arm_compute/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.h" +#include "arm_compute/runtime/CPP/functions/CPPDetectionOutputLayer.h" #include "arm_compute/runtime/CPP/functions/CPPPermute.h" #include "arm_compute/runtime/CPP/functions/CPPUpsample.h" diff --git a/arm_compute/runtime/CPP/functions/CPPDetectionOutputLayer.h b/arm_compute/runtime/CPP/functions/CPPDetectionOutputLayer.h new file mode 100644 index 0000000000..7f80948c81 --- /dev/null +++ b/arm_compute/runtime/CPP/functions/CPPDetectionOutputLayer.h @@ -0,0 +1,100 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_CPP_DETECTION_OUTPUT_LAYER_H__ +#define __ARM_COMPUTE_CPP_DETECTION_OUTPUT_LAYER_H__ + +#include "arm_compute/runtime/CPP/ICPPSimpleFunction.h" + +#include "arm_compute/core/Types.h" + +#include <map> + +namespace arm_compute +{ +class ITensor; + +// Normalized Bounding Box [xmin, ymin, xmax, ymax] +using NormalizedBBox = std::array<float, 4>; +// LabelBBox used for map label and bounding box +using LabelBBox = std::map<int, std::vector<NormalizedBBox>>; + +/** CPP Function to generate the detection output based on location and confidence + * predictions by doing non maximum suppression. + * + * @note Intended for use with MultiBox detection method. + */ +class CPPDetectionOutputLayer : public IFunction +{ +public: + /** Default constructor */ + CPPDetectionOutputLayer(); + /** Configure the detection output layer CPP kernel + * + * @param[in] input_loc The mbox location input tensor of size [C1, N]. Data types supported: F32. + * @param[in] input_conf The mbox confidence input tensor of size [C2, N]. Data types supported: F32. + * @param[in] input_priorbox The mbox prior box input tensor of size [C3, 2, N]. Data types supported: F32. + * @param[out] output The output tensor of size [7, M]. Data types supported: Same as @p input + * @param[in] info (Optional) DetectionOutputLayerInfo information. + * + * @note Output contains all the detections. Of those, only the ones selected by the valid region are valid. + */ + void configure(const ITensor *input_loc, const ITensor *input_conf, const ITensor *input_priorbox, ITensor *output, DetectionOutputLayerInfo info = DetectionOutputLayerInfo()); + /** Static function to check if given info will lead to a valid configuration of @ref CPPDetectionOutputLayer + * + * @param[in] input_loc The mbox location input tensor info. Data types supported: F32. + * @param[in] input_conf The mbox confidence input tensor info. Data types supported: F32. + * @param[in] input_priorbox The mbox prior box input tensor info. Data types supported: F32. + * @param[in] output The output tensor info. Data types supported: Same as @p input + * @param[in] info (Optional) DetectionOutputLayerInfo information. + * + * @return a status + */ + static Status validate(const ITensorInfo *input_loc, const ITensorInfo *input_conf, const ITensorInfo *input_priorbox, const ITensorInfo *output, + DetectionOutputLayerInfo info = DetectionOutputLayerInfo()); + // Inherited methods overridden: + void run() override; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CPPDetectionOutputLayer(const CPPDetectionOutputLayer &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CPPDetectionOutputLayer &operator=(const CPPDetectionOutputLayer &) = delete; + +private: + const ITensor *_input_loc; + const ITensor *_input_conf; + const ITensor *_input_priorbox; + ITensor *_output; + DetectionOutputLayerInfo _info; + + int _num_priors; + int _num; + + std::vector<LabelBBox> _all_location_predictions; + std::vector<std::map<int, std::vector<float>>> _all_confidence_scores; + std::vector<NormalizedBBox> _all_prior_bboxes; + std::vector<std::array<float, 4>> _all_prior_variances; + std::vector<LabelBBox> _all_decode_bboxes; + std::vector<std::map<int, std::vector<int>>> _all_indices; +}; +} // namespace arm_compute +#endif /* __ARM_COMPUTE_CPP_DETECTION_OUTPUT_LAYER_H__ */ diff --git a/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp b/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp new file mode 100644 index 0000000000..61005ab5fd --- /dev/null +++ b/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp @@ -0,0 +1,585 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/runtime/CPP/functions/CPPDetectionOutputLayer.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Validate.h" +#include "support/ToolchainSupport.h" + +#include <list> + +namespace arm_compute +{ +namespace +{ +Status validate_arguments(const ITensorInfo *input_loc, const ITensorInfo *input_conf, const ITensorInfo *input_priorbox, const ITensorInfo *output, DetectionOutputLayerInfo info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input_loc, input_conf, input_priorbox, output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_loc, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_loc, input_conf, input_priorbox); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_loc->num_dimensions() > 2, "The location input tensor should be [C1, N]."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_conf->num_dimensions() > 2, "The location input tensor should be [C2, N]."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_priorbox->num_dimensions() > 3, "The priorbox input tensor should be [C3, 2, N]."); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.eta() <= 0.f && info.eta() > 1.f, "Eta should be between 0 and 1"); + + const int num_priors = input_priorbox->tensor_shape()[0] / 4; + ARM_COMPUTE_RETURN_ERROR_ON_MSG(static_cast<size_t>((num_priors * info.num_loc_classes() * 4)) != input_loc->tensor_shape()[0], "Number of priors must match number of location predictions."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(static_cast<size_t>((num_priors * info.num_classes())) != input_conf->tensor_shape()[0], "Number of priors must match number of confidence predictions."); + + // Validate configured output + if(output->total_size() != 0) + { + const unsigned int max_size = info.keep_top_k() * (input_loc->num_dimensions() > 1 ? input_loc->dimension(1) : 1); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), TensorShape(7U, max_size)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_loc, output); + } + + return Status{}; +} + +/** Function used to sort pair<float, T> in descend order based on the score (first) value. + */ +template <typename T> +bool SortScorePairDescend(const std::pair<float, T> &pair1, + const std::pair<float, T> &pair2) +{ + return pair1.first > pair2.first; +} + +/** Get location predictions from input_loc. + * + * @param[in] input_loc The input location prediction. + * @param[in] num The number of images. + * @param[in] num_priors number of predictions per class. + * @param[in] num_loc_classes number of location classes. It is 1 if share_location is true, + * and is equal to number of classes needed to predict otherwise. + * @param[in] share_location If true, all classes share the same location prediction. + * @param[out] all_location_predictions All the location predictions. + * + */ +void retrieve_all_loc_predictions(const ITensor *input_loc, const int num, + const int num_priors, const int num_loc_classes, + const bool share_location, std::vector<LabelBBox> &all_location_predictions) +{ + for(int i = 0; i < num; ++i) + { + for(int c = 0; c < num_loc_classes; ++c) + { + int label = share_location ? -1 : c; + if(all_location_predictions[i].find(label) == all_location_predictions[i].end()) + { + all_location_predictions[i][label].resize(num_priors); + } + else + { + ARM_COMPUTE_ERROR_ON(all_location_predictions[i][label].size() != static_cast<size_t>(num_priors)); + break; + } + } + } + for(int i = 0; i < num; ++i) + { + for(int p = 0; p < num_priors; ++p) + { + for(int c = 0; c < num_loc_classes; ++c) + { + const int label = share_location ? -1 : c; + const int base_ptr = i * num_priors * num_loc_classes * 4 + p * num_loc_classes * 4 + c * 4; + //xmin, ymin, xmax, ymax + all_location_predictions[i][label][p][0] = *reinterpret_cast<float *>(input_loc->ptr_to_element(Coordinates(base_ptr))); + all_location_predictions[i][label][p][1] = *reinterpret_cast<float *>(input_loc->ptr_to_element(Coordinates(base_ptr + 1))); + all_location_predictions[i][label][p][2] = *reinterpret_cast<float *>(input_loc->ptr_to_element(Coordinates(base_ptr + 2))); + all_location_predictions[i][label][p][3] = *reinterpret_cast<float *>(input_loc->ptr_to_element(Coordinates(base_ptr + 3))); + } + } + } +} + +/** Get confidence predictions from input_conf. + * + * @param[in] input_loc The input location prediction. + * @param[in] num The number of images. + * @param[in] num_priors Number of predictions per class. + * @param[in] num_loc_classes Number of location classes. It is 1 if share_location is true, + * and is equal to number of classes needed to predict otherwise. + * @param[out] all_location_predictions All the location predictions. + * + */ +void retrieve_all_conf_scores(const ITensor *input_conf, const int num, + const int num_priors, const int num_classes, + std::vector<std::map<int, std::vector<float>>> &all_confidence_scores) +{ + std::vector<float> tmp_buffer; + tmp_buffer.resize(num * num_priors * num_classes); + for(int i = 0; i < num; ++i) + { + for(int c = 0; c < num_classes; ++c) + { + for(int p = 0; p < num_priors; ++p) + { + tmp_buffer[i * num_classes * num_priors + c * num_priors + p] = + *reinterpret_cast<float *>(input_conf->ptr_to_element(Coordinates(i * num_classes * num_priors + p * num_classes + c))); + } + } + } + for(int i = 0; i < num; ++i) + { + for(int c = 0; c < num_classes; ++c) + { + all_confidence_scores[i][c].resize(num_priors); + all_confidence_scores[i][c].assign(&tmp_buffer[i * num_classes * num_priors + c * num_priors], + &tmp_buffer[i * num_classes * num_priors + c * num_priors + num_priors]); + } + } +} + +/** Get prior boxes from input_priorbox. + * + * @param[in] input_priorbox The input location prediction. + * @param[in] num_priors Number of priors. + * @param[in] num_loc_classes number of location classes. It is 1 if share_location is true, + * and is equal to number of classes needed to predict otherwise. + * @param[out] all_prior_bboxes If true, all classes share the same location prediction. + * @param[out] all_location_predictions All the location predictions. + * + */ +void retrieve_all_priorbox(const ITensor *input_priorbox, + const int num_priors, + std::vector<NormalizedBBox> &all_prior_bboxes, + std::vector<std::array<float, 4>> &all_prior_variances) +{ + for(int i = 0; i < num_priors; ++i) + { + all_prior_bboxes[i] = { *reinterpret_cast<float *>(input_priorbox->ptr_to_element(Coordinates(i * 4))), + *reinterpret_cast<float *>(input_priorbox->ptr_to_element(Coordinates(i * 4 + 1))), + *reinterpret_cast<float *>(input_priorbox->ptr_to_element(Coordinates(i * 4 + 2))), + *reinterpret_cast<float *>(input_priorbox->ptr_to_element(Coordinates(i * 4 + 3))) + }; + } + + std::array<float, 4> var({ 0, 0, 0, 0 }); + for(int i = 0; i < num_priors; ++i) + { + for(int j = 0; j < 4; ++j) + { + var[j] = *reinterpret_cast<float *>(input_priorbox->ptr_to_element(Coordinates((num_priors + i) * 4 + j))); + } + all_prior_variances[i] = var; + } +} + +/** Decode a bbox according to a prior bbox. + * + * @param[in] prior_bbox The input prior bounding boxes. + * @param[in] prior_variance The corresponding input variance. + * @param[in] code_type The detection output code type used to decode the results. + * @param[in] variance_encoded_in_target If true, the variance is encoded in target. + * @param[in] clip_bbox If true, the results should be between 0.f and 1.f. + * @param[in] bbox The input bbox to decode + * @param[out] decode_bbox The decoded bboxes. + * + */ +void DecodeBBox(const NormalizedBBox &prior_bbox, const std::array<float, 4> &prior_variance, + const DetectionOutputLayerCodeType code_type, const bool variance_encoded_in_target, + const bool clip_bbox, const NormalizedBBox &bbox, NormalizedBBox &decode_bbox) +{ + // if the variance is encoded in target, we simply need to add the offset predictions + // otherwise we need to scale the offset accordingly. + switch(code_type) + { + case DetectionOutputLayerCodeType::CORNER: + { + decode_bbox[0] = prior_bbox[0] + (variance_encoded_in_target ? bbox[0] : prior_variance[0] * bbox[0]); + decode_bbox[1] = prior_bbox[1] + (variance_encoded_in_target ? bbox[1] : prior_variance[1] * bbox[1]); + decode_bbox[2] = prior_bbox[2] + (variance_encoded_in_target ? bbox[2] : prior_variance[2] * bbox[2]); + decode_bbox[3] = prior_bbox[3] + (variance_encoded_in_target ? bbox[3] : prior_variance[3] * bbox[3]); + + break; + } + case DetectionOutputLayerCodeType::CENTER_SIZE: + { + const float prior_width = prior_bbox[2] - prior_bbox[0]; + const float prior_height = prior_bbox[3] - prior_bbox[1]; + + // Check if the prior width and height are right + ARM_COMPUTE_ERROR_ON(prior_width <= 0.f); + ARM_COMPUTE_ERROR_ON(prior_height <= 0.f); + + const float prior_center_x = (prior_bbox[0] + prior_bbox[2]) / 2.; + const float prior_center_y = (prior_bbox[1] + prior_bbox[3]) / 2.; + + const float decode_bbox_center_x = (variance_encoded_in_target ? bbox[0] : prior_variance[0] * bbox[0]) * prior_width + prior_center_x; + const float decode_bbox_center_y = (variance_encoded_in_target ? bbox[1] : prior_variance[1] * bbox[1]) * prior_height + prior_center_y; + const float decode_bbox_width = (variance_encoded_in_target ? std::exp(bbox[2]) : std::exp(prior_variance[2] * bbox[2])) * prior_width; + const float decode_bbox_height = (variance_encoded_in_target ? std::exp(bbox[3]) : std::exp(prior_variance[3] * bbox[3])) * prior_height; + + decode_bbox[0] = (decode_bbox_center_x - decode_bbox_width / 2.f); + decode_bbox[1] = (decode_bbox_center_y - decode_bbox_height / 2.f); + decode_bbox[2] = (decode_bbox_center_x + decode_bbox_width / 2.f); + decode_bbox[3] = (decode_bbox_center_y + decode_bbox_height / 2.f); + + break; + } + case DetectionOutputLayerCodeType::CORNER_SIZE: + { + const float prior_width = prior_bbox[2] - prior_bbox[0]; + const float prior_height = prior_bbox[3] - prior_bbox[1]; + + // Check if the prior width and height are greater than 0 + ARM_COMPUTE_ERROR_ON(prior_width <= 0.f); + ARM_COMPUTE_ERROR_ON(prior_height <= 0.f); + + decode_bbox[0] = prior_bbox[0] + (variance_encoded_in_target ? bbox[0] : prior_variance[0] * bbox[0]) * prior_width; + decode_bbox[1] = prior_bbox[1] + (variance_encoded_in_target ? bbox[1] : prior_variance[1] * bbox[1]) * prior_height; + decode_bbox[2] = prior_bbox[2] + (variance_encoded_in_target ? bbox[2] : prior_variance[2] * bbox[2]) * prior_width; + decode_bbox[3] = prior_bbox[3] + (variance_encoded_in_target ? bbox[3] : prior_variance[3] * bbox[3]) * prior_height; + + break; + } + default: + ARM_COMPUTE_ERROR("Unsupported Detection Output Code Type."); + } + + if(clip_bbox) + { + for(auto &d_bbox : decode_bbox) + { + d_bbox = utility::clamp(d_bbox, 0.f, 1.f); + } + } +} + +/** Do non maximum suppression given bboxes and scores. + * + * @param[in] bboxes The input bounding boxes. + * @param[in] scores The corresponding input confidence. + * @param[in] score_threshold The threshold used to filter detection results. + * @param[in] nms_threshold The threshold used in non maximum suppression. + * @param[in] eta Adaptation rate for nms threshold. + * @param[in] top_k If not -1, keep at most top_k picked indices. + * @param[out] indices The kept indices of bboxes after nms. + * + */ +void ApplyNMSFast(const std::vector<NormalizedBBox> &bboxes, + const std::vector<float> &scores, const float score_threshold, + const float nms_threshold, const float eta, const int top_k, + std::vector<int> &indices) +{ + ARM_COMPUTE_ERROR_ON_MSG(bboxes.size() != scores.size(), "bboxes and scores have different size."); + + // Get top_k scores (with corresponding indices). + std::list<std::pair<float, int>> score_index_vec; + + // Generate index score pairs. + for(size_t i = 0; i < scores.size(); ++i) + { + if(scores[i] > score_threshold) + { + score_index_vec.emplace_back(std::make_pair(scores[i], i)); + } + } + + // Sort the score pair according to the scores in descending order + score_index_vec.sort(SortScorePairDescend<int>); + + // Keep top_k scores if needed. + const int score_index_vec_size = score_index_vec.size(); + if(top_k > -1 && top_k < score_index_vec_size) + { + score_index_vec.resize(top_k); + } + + // Do nms. + float adaptive_threshold = nms_threshold; + indices.clear(); + + while(!score_index_vec.empty()) + { + const int idx = score_index_vec.front().second; + bool keep = true; + for(int kept_idx : indices) + { + if(keep) + { + // Compute the jaccard (intersection over union IoU) overlap between two bboxes. + NormalizedBBox intersect_bbox = std::array<float, 4>({ 0, 0, 0, 0 }); + if(bboxes[kept_idx][0] > bboxes[idx][2] || bboxes[kept_idx][2] < bboxes[idx][0] || bboxes[kept_idx][1] > bboxes[idx][3] || bboxes[kept_idx][3] < bboxes[idx][1]) + { + intersect_bbox = std::array<float, 4>({ 0, 0, 0, 0 }); + } + else + { + intersect_bbox = std::array<float, 4>({ std::max(bboxes[idx][0], bboxes[kept_idx][0]), std::max(bboxes[idx][1], bboxes[kept_idx][1]), std::min(bboxes[idx][2], bboxes[kept_idx][2]), std::min(bboxes[idx][3], + bboxes[kept_idx][3]) + }); + } + + float intersect_width = intersect_bbox[2] - intersect_bbox[0]; + float intersect_height = intersect_bbox[3] - intersect_bbox[1]; + + float overlap = 0.f; + if(intersect_width > 0 && intersect_height > 0) + { + float intersect_size = intersect_width * intersect_height; + float bbox1_size = (bboxes[idx][2] < bboxes[idx][0] + || bboxes[idx][3] < bboxes[idx][1]) ? + 0.f : + (bboxes[idx][2] - bboxes[idx][0]) * (bboxes[idx][3] - bboxes[idx][1]); //BBoxSize(bboxes[idx]); + float bbox2_size = (bboxes[kept_idx][2] < bboxes[kept_idx][0] + || bboxes[kept_idx][3] < bboxes[kept_idx][1]) ? + 0.f : + (bboxes[kept_idx][2] - bboxes[kept_idx][0]) * (bboxes[kept_idx][3] - bboxes[kept_idx][1]); // BBoxSize(bboxes[kept_idx]); + overlap = intersect_size / (bbox1_size + bbox2_size - intersect_size); + } + keep = (overlap <= adaptive_threshold); + } + else + { + break; + } + } + if(keep) + { + indices.push_back(idx); + } + score_index_vec.erase(score_index_vec.begin()); + if(keep && eta < 1 && adaptive_threshold > 0.5) + { + adaptive_threshold *= eta; + } + } +} +} // namespace + +CPPDetectionOutputLayer::CPPDetectionOutputLayer() + : _input_loc(nullptr), _input_conf(nullptr), _input_priorbox(nullptr), _output(nullptr), _info(), _num_priors(), _num(), _all_location_predictions(), _all_confidence_scores(), _all_prior_bboxes(), + _all_prior_variances(), _all_decode_bboxes(), _all_indices() +{ +} + +void CPPDetectionOutputLayer::configure(const ITensor *input_loc, const ITensor *input_conf, const ITensor *input_priorbox, ITensor *output, DetectionOutputLayerInfo info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input_loc, input_conf, input_priorbox, output); + // Output auto initialization if not yet initialized + // Since the number of bboxes to kept is unknown before nms, the shape is set to the maximum + // The maximum is keep_top_k * input_loc_size[1] + // Each row is a 7 dimension std::vector, which stores [image_id, label, confidence, xmin, ymin, xmax, ymax] + const unsigned int max_size = info.keep_top_k() * (input_loc->info()->num_dimensions() > 1 ? input_loc->info()->dimension(1) : 1); + auto_init_if_empty(*output->info(), input_loc->info()->clone()->set_tensor_shape(TensorShape(7U, max_size))); + + // Perform validation step + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input_loc->info(), input_conf->info(), input_priorbox->info(), output->info(), info)); + + _input_loc = input_loc; + _input_conf = input_conf; + _input_priorbox = input_priorbox; + _output = output; + _info = info; + _num_priors = input_priorbox->info()->dimension(0) / 4; + _num = (_input_loc->info()->num_dimensions() > 1 ? _input_loc->info()->dimension(1) : 1); + + _all_location_predictions.resize(_num); + _all_confidence_scores.resize(_num); + _all_prior_bboxes.resize(_num_priors); + _all_prior_variances.resize(_num_priors); + _all_decode_bboxes.resize(_num); + + for(int i = 0; i < _num; ++i) + { + for(int c = 0; c < _info.num_loc_classes(); ++c) + { + const int label = _info.share_location() ? -1 : c; + if(label == _info.background_label_id()) + { + // Ignore background class. + continue; + } + _all_decode_bboxes[i][label].resize(_num_priors); + } + } + _all_indices.resize(_num); + + Coordinates coord; + coord.set_num_dimensions(output->info()->num_dimensions()); + output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape())); +} + +Status CPPDetectionOutputLayer::validate(const ITensorInfo *input_loc, const ITensorInfo *input_conf, const ITensorInfo *input_priorbox, const ITensorInfo *output, DetectionOutputLayerInfo info) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input_loc, input_conf, input_priorbox, output, info)); + return Status{}; +} + +void CPPDetectionOutputLayer::run() +{ + // Retrieve all location predictions. + retrieve_all_loc_predictions(_input_loc, _num, _num_priors, _info.num_loc_classes(), _info.share_location(), _all_location_predictions); + + // Retrieve all confidences. + retrieve_all_conf_scores(_input_conf, _num, _num_priors, _info.num_classes(), _all_confidence_scores); + + // Retrieve all prior bboxes. + retrieve_all_priorbox(_input_priorbox, _num_priors, _all_prior_bboxes, _all_prior_variances); + + // Decode all loc predictions to bboxes + const bool clip_bbox = false; + for(int i = 0; i < _num; ++i) + { + for(int c = 0; c < _info.num_loc_classes(); ++c) + { + const int label = _info.share_location() ? -1 : c; + if(label == _info.background_label_id()) + { + // Ignore background class. + continue; + } + ARM_COMPUTE_ERROR_ON_MSG(_all_location_predictions[i].find(label) == _all_location_predictions[i].end(), "Could not find location predictions for label %d.", label); + + const std::vector<NormalizedBBox> &label_loc_preds = _all_location_predictions[i].find(label)->second; + + const int num_bboxes = _all_prior_bboxes.size(); + ARM_COMPUTE_ERROR_ON(_all_prior_variances[i].size() != 4); + + for(int j = 0; j < num_bboxes; ++j) + { + DecodeBBox(_all_prior_bboxes[j], _all_prior_variances[j], _info.code_type(), _info.variance_encoded_in_target(), clip_bbox, label_loc_preds[j], _all_decode_bboxes[i][label][j]); + } + } + } + + int num_kept = 0; + + for(int i = 0; i < _num; ++i) + { + const LabelBBox &decode_bboxes = _all_decode_bboxes[i]; + const std::map<int, std::vector<float>> &conf_scores = _all_confidence_scores[i]; + + std::map<int, std::vector<int>> indices; + int num_det = 0; + for(int c = 0; c < _info.num_classes(); ++c) + { + if(c == _info.background_label_id()) + { + // Ignore background class + continue; + } + const int label = _info.share_location() ? -1 : c; + if(conf_scores.find(c) == conf_scores.end() || decode_bboxes.find(label) == decode_bboxes.end()) + { + ARM_COMPUTE_ERROR("Could not find predictions for label %d.", label); + } + const std::vector<float> &scores = conf_scores.find(c)->second; + const std::vector<NormalizedBBox> &bboxes = decode_bboxes.find(label)->second; + + ApplyNMSFast(bboxes, scores, _info.confidence_threshold(), _info.nms_threshold(), _info.eta(), _info.top_k(), indices[c]); + + num_det += indices[c].size(); + } + + int num_to_add = 0; + if(_info.keep_top_k() > -1 && num_det > _info.keep_top_k()) + { + std::vector<std::pair<float, std::pair<int, int>>> score_index_pairs; + for(auto it : indices) + { + const int label = it.first; + const std::vector<int> &label_indices = it.second; + + if(conf_scores.find(label) == conf_scores.end()) + { + ARM_COMPUTE_ERROR("Could not find predictions for label %d.", label); + } + + const std::vector<float> &scores = conf_scores.find(label)->second; + for(auto idx : label_indices) + { + ARM_COMPUTE_ERROR_ON(idx > static_cast<int>(scores.size())); + score_index_pairs.push_back(std::make_pair(scores[idx], std::make_pair(label, idx))); + } + } + + // Keep top k results per image. + std::sort(score_index_pairs.begin(), score_index_pairs.end(), SortScorePairDescend<std::pair<int, int>>); + score_index_pairs.resize(_info.keep_top_k()); + + // Store the new indices. + + std::map<int, std::vector<int>> new_indices; + for(auto score_index_pair : score_index_pairs) + { + int label = score_index_pair.second.first; + int idx = score_index_pair.second.second; + new_indices[label].push_back(idx); + } + _all_indices[i] = new_indices; + num_to_add = _info.keep_top_k(); + } + else + { + _all_indices[i] = indices; + num_to_add = num_det; + } + num_kept += num_to_add; + } + + //Update the valid region of the ouput to mark the exact number of detection + _output->info()->set_valid_region(ValidRegion(Coordinates(0, 0), TensorShape(7, num_kept))); + + int count = 0; + for(int i = 0; i < _num; ++i) + { + const std::map<int, std::vector<float>> &conf_scores = _all_confidence_scores[i]; + const LabelBBox &decode_bboxes = _all_decode_bboxes[i]; + for(auto &it : _all_indices[i]) + { + const int label = it.first; + const std::vector<float> &scores = conf_scores.find(label)->second; + const int loc_label = _info.share_location() ? -1 : label; + if(conf_scores.find(label) == conf_scores.end() || decode_bboxes.find(loc_label) == decode_bboxes.end()) + { + // Either if there are no confidence predictions + // or there are no location predictions for current label. + ARM_COMPUTE_ERROR("Could not find predictions for the label %d.", label); + } + const std::vector<NormalizedBBox> &bboxes = decode_bboxes.find(loc_label)->second; + const std::vector<int> &indices = it.second; + + for(auto idx : indices) + { + *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(count * 7)))) = i; + *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(count * 7 + 1)))) = label; + *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(count * 7 + 2)))) = scores[idx]; + *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(count * 7 + 3)))) = bboxes[idx][0]; + *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(count * 7 + 4)))) = bboxes[idx][1]; + *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(count * 7 + 5)))) = bboxes[idx][2]; + *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(count * 7 + 6)))) = bboxes[idx][3]; + + ++count; + } + } + } +} +} // namespace arm_compute
\ No newline at end of file diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h index 6850ea018a..496e49beb1 100644 --- a/utils/TypePrinter.h +++ b/utils/TypePrinter.h @@ -1901,6 +1901,85 @@ inline ::std::ostream &operator<<(::std::ostream &os, const DetectionWindow &det return os; } +/** Formatted output of the DetectionOutputLayerCodeType type. + * + * @param[out] os Output stream + * @param[in] detection_code Type to output + * + * @return Modified output stream. + */ +inline ::std::ostream &operator<<(::std::ostream &os, const DetectionOutputLayerCodeType &detection_code) +{ + switch(detection_code) + { + case DetectionOutputLayerCodeType::CENTER_SIZE: + os << "CENTER_SIZE"; + break; + case DetectionOutputLayerCodeType::CORNER: + os << "CORNER"; + break; + case DetectionOutputLayerCodeType::CORNER_SIZE: + os << "CORNER_SIZE"; + break; + case DetectionOutputLayerCodeType::TF_CENTER: + os << "TF_CENTER"; + break; + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } + + return os; +} +/** Formatted output of the DetectionOutputLayerCodeType type. + * + * @param[in] detection_code Type to output + * + * @return Formatted string. + */ +inline std::string to_string(const DetectionOutputLayerCodeType &detection_code) +{ + std::stringstream str; + str << detection_code; + return str.str(); +} + +/** Formatted output of the DetectionOutputLayerInfo type. + * + * @param[out] os Output stream + * @param[in] detection_info Type to output + * + * @return Modified output stream. + */ +inline ::std::ostream &operator<<(::std::ostream &os, const DetectionOutputLayerInfo &detection_info) +{ + os << "{Classes=" << detection_info.num_classes() << "," + << "ShareLocation=" << detection_info.share_location() << "," + << "CodeType=" << detection_info.code_type() << "," + << "VarianceEncodedInTarget=" << detection_info.variance_encoded_in_target() << "," + << "KeepTopK=" << detection_info.keep_top_k() << "," + << "NMSThreshold=" << detection_info.nms_threshold() << "," + << "Eta=" << detection_info.eta() << "," + << "BackgroundLabelId=" << detection_info.background_label_id() << "," + << "ConfidenceThreshold=" << detection_info.confidence_threshold() << "," + << "TopK=" << detection_info.top_k() << "," + << "NumLocClasses=" << detection_info.num_loc_classes() + << "}"; + + return os; +} + +/** Formatted output of the DetectionOutputLayerInfo type. + * + * @param[in] detection_info Type to output + * + * @return Formatted string. + */ +inline std::string to_string(const DetectionOutputLayerInfo &detection_info) +{ + std::stringstream str; + str << detection_info; + return str.str(); +} /** Formatted output of the DetectionWindow type. * * @param[in] detection_window Type to output |