prediction_dict = {}
rpn_feature = self._rpn_activation(self._rpn(conv_feature_map))
// Then we apply separate conv layers for classification and regression.
rpn_cls_score_original = self._rpn_cls(rpn_feature)
rpn_bbox_pred_original = self._rpn_bbox(rpn_feature)
// rpn_cls_score_original has shape (1, H, W, num_anchors * 2)
// rpn_bbox_pred_original has shape (1, H, W, num_anchors * 4)
// where H, W are height and width of the pretrained feature map.
// Convert (flatten) `rpn_cls_score_original` which has two scalars per
// anchor per location to be able to apply softmax.
rpn_cls_score = tf.reshape(rpn_cls_score_original, [-1, 2])
// Now that `rpn_cls_score` has shape (H * W * num_anchors, 2), we apply
// softmax to the last dim.
rpn_cls_prob = tf.nn.softmax(rpn_cls_score)
prediction_dict["rpn_cls_prob"] = rpn_cls_prob
prediction_dict["rpn_cls_score"] = rpn_cls_score
// Flatten bounding box delta prediction for easy manipulation.
// We end up with `rpn_bbox_pred` having shape (H * W * num_anchors, 4).
rpn_bbox_pred = tf.reshape(rpn_bbox_pred_original, [-1, 4])
prediction_dict["rpn_bbox_pred"] = rpn_bbox_pred
// We have to convert bbox deltas to usable bounding boxes and remove
// redundant ones using Non Maximum Suppression (NMS).
proposal_prediction = self._proposal(
rpn_cls_prob, rpn_bbox_pred, all_anchors, im_shape)
prediction_dict["proposals"] = proposal_prediction["nms_proposals"]
prediction_dict["scores"] = proposal_prediction["nms_proposals_scores"]
if self._debug:
prediction_dict["proposal_prediction"] = proposal_prediction
if gt_boxes is not None:
// When training we use a separate module to calculate the target
After Change
prediction_dict = {}
rpn_feature = self._rpn(conv_feature_map)
// Then we apply separate conv layers for classification and regression.
rpn_cls_score_original = self._rpn_cls(rpn_feature)
rpn_bbox_pred_original = self._rpn_bbox(rpn_feature)
// rpn_cls_score_original has shape (1, H, W, num_anchors * 2)
// rpn_bbox_pred_original has shape (1, H, W, num_anchors * 4)
// where H, W are height and width of the pretrained feature map.
// Convert (flatten) `rpn_cls_score_original` which has two scalars per
// anchor per location to be able to apply softmax.
rpn_cls_score = tf.reshape(rpn_cls_score_original, [-1, 2])
// Now that `rpn_cls_score` has shape (H * W * num_anchors, 2), we apply
// softmax to the last dim.
rpn_cls_prob = tf.nn.softmax(rpn_cls_score)
prediction_dict["rpn_cls_prob"] = rpn_cls_prob
prediction_dict["rpn_cls_score"] = rpn_cls_score
// Flatten bounding box delta prediction for easy manipulation.
// We end up with `rpn_bbox_pred` having shape (H * W * num_anchors, 4).
rpn_bbox_pred = tf.reshape(rpn_bbox_pred_original, [-1, 4])
prediction_dict["rpn_bbox_pred"] = rpn_bbox_pred
// We have to convert bbox deltas to usable bounding boxes and remove
// redundant ones using Non Maximum Suppression (NMS).
proposal_prediction = self._proposal(
rpn_cls_prob, rpn_bbox_pred, all_anchors, im_shape)
prediction_dict["proposals"] = proposal_prediction["nms_proposals"]
prediction_dict["scores"] = proposal_prediction["nms_proposals_scores"]
if self._debug:
prediction_dict["proposal_prediction"] = proposal_prediction
if gt_boxes is not None:
// When training we use a separate module to calculate the target
