diff --git a/scripts/libraries/ml/ml/postprocessing.py b/scripts/libraries/ml/ml/postprocessing.py index cd6b833fa..badcc03b5 100644 --- a/scripts/libraries/ml/ml/postprocessing.py +++ b/scripts/libraries/ml/ml/postprocessing.py @@ -64,6 +64,8 @@ class fomo_postprocess: return nms.get_bounding_boxes() +# This is a lightweight version of the tiny yolo v2 object detection algorithm. +# It was optimized to work well on embedded devices with limited computational resources. class yolo_v2_postprocess: _YOLO_V2_TX = const(0) _YOLO_V2_TY = const(1) @@ -100,19 +102,17 @@ class yolo_v2_postprocess: return e_x / np.sum(e_x, axis=1, keepdims=True) # Reshape the output to a 2D array - colum_outputs = outputs[0].reshape((oh * ow * self.anchors_len, - _YOLO_V2_CLASSES + class_count)) + row_outputs = outputs[0].reshape((oh * ow * self.anchors_len, + _YOLO_V2_CLASSES + class_count)) # Threshold all the scores - score_indices = sigmoid(colum_outputs[:, _YOLO_V2_SCORE]) - score_indices = np.nonzero(score_indices > self.threshold) - if isinstance(score_indices, tuple): - score_indices = score_indices[0] + score_indices = sigmoid(row_outputs[:, _YOLO_V2_SCORE]) + score_indices = np.nonzero(score_indices > self.threshold)[0] if not len(score_indices): return _NO_DETECTION # Get the bounding boxes that have a valid score - bb = np.take(colum_outputs, score_indices, axis=0) + bb = np.take(row_outputs, score_indices, axis=0) # Extract rows, columns, and anchor indices bb_rows = score_indices // (ow * self.anchors_len) @@ -142,7 +142,7 @@ class yolo_v2_postprocess: h_rel = h_rel * ih nms = NMS(iw, ih, inputs[0].roi) - for i in range(len(bb)): + for i in range(bb.shape[0]): nms.add_bounding_box(x_center[i] - (w_rel[i] / 2), y_center[i] - (h_rel[i] / 2), x_center[i] + (w_rel[i] / 2), @@ -151,6 +151,19 @@ class yolo_v2_postprocess: return nms.get_bounding_boxes(threshold=self.nms_threshold, sigma=self.nms_sigma) +# This is a lightweight version of the YOLO (You Only Look Once) object detection algorithm. +# It is designed to work well on embedded devices with limited computational resources. +class yolo_lc_postprocess(yolo_v2_postprocess): + def __init__(self, threshold=0.6, anchors=None, nms_threshold=0.1, nms_sigma=0.1): + if anchors is None: + anchors = np.array([[0.076023, 0.258508], + [0.163031, 0.413531], + [0.234769, 0.702585], + [0.427054, 0.715892], + [0.748154, 0.857092]]) + super().__init__(threshold, anchors, nms_threshold, nms_sigma) + + class yolo_v5_postprocess: _YOLO_V5_CX = const(0) _YOLO_V5_CY = const(1) @@ -169,18 +182,16 @@ class yolo_v5_postprocess: class_count = oc - _YOLO_V5_CLASSES # Reshape the output to a 2D array - colum_outputs = outputs[0].reshape((oh * ow, _YOLO_V5_CLASSES + class_count)) + row_outputs = outputs[0].reshape((oh * ow, _YOLO_V5_CLASSES + class_count)) # Threshold all the scores - score_indices = colum_outputs[:, _YOLO_V5_SCORE] - score_indices = np.nonzero(score_indices > self.threshold) - if isinstance(score_indices, tuple): - score_indices = score_indices[0] + score_indices = row_outputs[:, _YOLO_V5_SCORE] + score_indices = np.nonzero(score_indices > self.threshold)[0] if not len(score_indices): return _NO_DETECTION # Get the bounding boxes that have a valid score - bb = np.take(colum_outputs, score_indices, axis=0) + bb = np.take(row_outputs, score_indices, axis=0) # Get the score information bb_scores = bb[:, _YOLO_V5_SCORE] @@ -202,7 +213,61 @@ class yolo_v5_postprocess: ymax = (y_center + h_rel) * ih nms = NMS(iw, ih, inputs[0].roi) - for i in range(len(bb)): + for i in range(bb.shape[0]): + nms.add_bounding_box(xmin[i], ymin[i], xmax[i], ymax[i], + bb_scores[i], bb_classes[i]) + return nms.get_bounding_boxes(threshold=self.nms_threshold, sigma=self.nms_sigma) + + +class yolo_v8_postprocess: + _YOLO_V8_CX = const(0) + _YOLO_V8_CY = const(1) + _YOLO_V8_CW = const(2) + _YOLO_V8_CH = const(3) + _YOLO_V8_CLASSES = const(4) + + def __init__(self, threshold=0.6, nms_threshold=0.1, nms_sigma=0.1): + self.threshold = threshold + self.nms_threshold = nms_threshold + self.nms_sigma = nms_sigma + + def __call__(self, model, inputs, outputs): + oh, ow, oc = model.output_shape[0] + class_count = ow - _YOLO_V8_CLASSES + + # Reshape the output to a 2D array + column_outputs = outputs[0].reshape((oh * (_YOLO_V8_CLASSES + class_count), oc)) + + # Threshold all the scores + score_indices = np.max(column_outputs[_YOLO_V8_CLASSES:, :], axis=0) + score_indices = np.nonzero(score_indices > self.threshold)[0] + if not len(score_indices): + return _NO_DETECTION + + # Get the bounding boxes that have a valid score + bb = np.take(column_outputs, score_indices, axis=1) + + # Get the score information + bb_scores = np.max(bb[_YOLO_V8_CLASSES:, :], axis=0) + + # Get the class information + bb_classes = np.argmax(bb[_YOLO_V8_CLASSES:, :], axis=0) + + # Compute the bounding box information + x_center = bb[_YOLO_V8_CX, :] + y_center = bb[_YOLO_V8_CY, :] + w_rel = bb[_YOLO_V8_CW, :] * 0.5 + h_rel = bb[_YOLO_V8_CH, :] * 0.5 + + # Scale the bounding boxes to have enough integer precision for NMS + ib, ih, iw, ic = model.input_shape[0] + xmin = (x_center - w_rel) * iw + ymin = (y_center - h_rel) * ih + xmax = (x_center + w_rel) * iw + ymax = (y_center + h_rel) * ih + + nms = NMS(iw, ih, inputs[0].roi) + for i in range(bb.shape[1]): nms.add_bounding_box(xmin[i], ymin[i], xmax[i], ymax[i], bb_scores[i], bb_classes[i]) return nms.get_bounding_boxes(threshold=self.nms_threshold, sigma=self.nms_sigma)