|
| 1 | +#include "yolov3.h" |
| 2 | + |
| 3 | +std::vector<std::vector<float>> YOLOV3::tensorToVector2D() { |
| 4 | + std::vector<std::vector<float>> v; |
| 5 | + if (_interpreter->outputs().size() != 3) { |
| 6 | + std::cout << "yolov3 don't support this model!\n"; |
| 7 | + std::cout << __FILE__ << ": " << __LINE__ << std::endl; |
| 8 | + exit(-1); |
| 9 | + } |
| 10 | + const size_t num_anchors = 3; |
| 11 | + int masks[3][3] = { |
| 12 | + {6, 7, 8}, |
| 13 | + {3, 4, 5}, |
| 14 | + {0, 1, 2}, |
| 15 | + }; |
| 16 | + float anchors[18] = {10, 13, 16, 30, 33, 23, 30, 61, 62, |
| 17 | + 45, 59, 119, 116, 90, 156, 198, 373, 326}; |
| 18 | + |
| 19 | + for (size_t grid = 0; grid < _interpreter->outputs().size(); grid++) { |
| 20 | + int out = _interpreter->outputs()[grid]; |
| 21 | + TfLiteIntArray *out_dims = _interpreter->tensor(out)->dims; |
| 22 | + TfLiteTensor *pOutputTensor = _interpreter->tensor(out); |
| 23 | + int out_batch = out_dims->data[0]; |
| 24 | + _out_row = out_dims->data[1]; |
| 25 | + _out_colum = out_dims->data[2]; |
| 26 | + _out_channel = out_dims->data[3]; |
| 27 | + std::cout << "GRID Output Shape:[" << out_batch << "][" << _out_row << "][" |
| 28 | + << _out_colum << "][" << _out_channel << "]\n"; |
| 29 | + |
| 30 | + if (pOutputTensor->type == kTfLiteFloat32) { |
| 31 | + for (size_t i = 0; i < _out_row; i++) { |
| 32 | + for (size_t j = 0; j < _out_colum; j++) { |
| 33 | + for (size_t k = 0; k < num_anchors; k++) { |
| 34 | + std::vector<float> vtem; |
| 35 | + for (int l = 0; l < _out_channel / num_anchors; l++) { |
| 36 | + float val_float = |
| 37 | + pOutputTensor->data |
| 38 | + .f[i * _out_colum * _out_channel + j * _out_channel + |
| 39 | + k * _out_channel / num_anchors + l]; |
| 40 | + if (l != 2 && l != 3) { |
| 41 | + val_float = 1. / (1. + exp(-val_float)); // logistic |
| 42 | + } |
| 43 | + vtem.push_back(val_float); |
| 44 | + } |
| 45 | + vtem[0] = (j + vtem[0]) / _out_colum; |
| 46 | + vtem[1] = (i + vtem[1]) / _out_row; |
| 47 | + vtem[2] = exp(vtem[2]) * anchors[2 * masks[grid][k]] / _in_width; |
| 48 | + vtem[3] = |
| 49 | + exp(vtem[3]) * anchors[2 * masks[grid][k] + 1] / _in_height; |
| 50 | + v.push_back(vtem); |
| 51 | + } |
| 52 | + } |
| 53 | + } |
| 54 | + } else { |
| 55 | + std::cout << "Unsupported output type!\n"; |
| 56 | + std::cout << __FILE__ << ": " << __LINE__ << std::endl; |
| 57 | + exit(-1); |
| 58 | + } |
| 59 | + } |
| 60 | + return v; |
| 61 | +} |
| 62 | + |
| 63 | +void YOLOV3::nonMaximumSupprition(std::vector<std::vector<float>> &predV, |
| 64 | + std::vector<cv::Rect> &boxes, |
| 65 | + std::vector<float> &confidences, |
| 66 | + std::vector<int> &classIds, |
| 67 | + std::vector<int> &indices) |
| 68 | + |
| 69 | +{ |
| 70 | + std::vector<cv::Rect> boxesNMS; |
| 71 | + std::vector<float> scores; |
| 72 | + double confidence; |
| 73 | + cv::Point classId; |
| 74 | + for (int i = 0; i < predV.size(); i++) { |
| 75 | + if (predV[i][4] > _conf_threshold) { |
| 76 | + int left = (predV[i][0] - predV[i][2] / 2) * _img_width; |
| 77 | + int top = (predV[i][1] - predV[i][3] / 2) * _img_height; |
| 78 | + int w = predV[i][2] * _img_width; |
| 79 | + int h = predV[i][3] * _img_height; |
| 80 | + |
| 81 | + for (int j = 5; j < 85; j++) { |
| 82 | + // # conf = obj_conf * cls_conf |
| 83 | + scores.push_back(predV[i][j] * predV[i][4]); |
| 84 | + } |
| 85 | + |
| 86 | + cv::minMaxLoc(scores, 0, &confidence, 0, &classId); |
| 87 | + |
| 88 | + scores.clear(); |
| 89 | + |
| 90 | + if (confidence > _conf_threshold * _conf_threshold) { |
| 91 | + boxes.push_back(cv::Rect(left, top, w, h)); |
| 92 | + confidences.push_back(confidence); |
| 93 | + classIds.push_back(classId.x); |
| 94 | + boxesNMS.push_back(cv::Rect(left, top, w, h)); |
| 95 | + } |
| 96 | + } |
| 97 | + } |
| 98 | + cv::dnn::NMSBoxes(boxesNMS, confidences, _conf_threshold, _nms_threshold, |
| 99 | + indices); |
| 100 | +} |
0 commit comments