0212b83a3ddf2ec1d217647673454626c386717a,frcnn/rcnn_target.py,RCNNTarget,proposal_target_layer,#RCNNTarget#Any#Any#Any#Any#,38

Before Change

        // We have to substract 1 because of the offset of the background class.
        proposals_label_with_target = proposals_label[proposal_with_target_idx] - 1
        // We cast them as ints since we want to use it for indexing.
        proposals_label_with_target = proposals_label_with_target.astype(np.int32)
        // We create a numpy array with shape (num_positive_proposals, 4),
        // having the ideal GT boxes for each proposal.
        proposals_gt_boxes = gt_boxes[proposals_label_with_target]
        // We create the same array but with the proposals

After Change

        proposal_with_target_idx = np.nonzero(proposals_label > 0)[0]

        // Get top gt_box for every proposal, top_gt_idx shape (1000,) with values < gt_boxes.shape[0]
        top_gt_idx = overlaps.argmax(axis=1)

        // Get the corresponding ground truth box only for the proposals with target.
        gt_boxes_ids = top_gt_idx[proposal_with_target_idx]

        // Get the values of the ground truth boxes. This is shaped (num_proposals, 5) because we also have the label.
        proposals_gt_boxes = gt_boxes[gt_boxes_ids]

Instances