目标检测 pytorch复现Yolov2目标检测项目

- YOVLOV2改进：
   - 1. 用Kmeans获取先验框的宽高
   - 2. 引入anchor。BS*7*7*2Boxes*30-->BS*13*13*5Boxes*(4+1+cls)
   - 3. 提高输入数据分辨率。224-->448
   - 4. 采用darknet-19主干网络。
   - 5. 特征融合(passthrough)。BS*26*26*512-->BS*13*13*256-->cat BS*13*13*1024-->BS*13*13*1280
   - 6. 加入BN。去掉fc、dropout，添加bn+relu,提高2%mAP。
   - 7. 预测中心点坐标偏移。解码函数改变。
   - 8. 多尺度输入训练网络。[320,352,384,448,480,512,544,576,608] 10/pre epoch
   

YOLOv2损失函数loss

代码：

def yolo_loss(args,
              anchors,
              num_classes,
              rescore_confidence=False,
              print_loss=False):
    """YOLO localization loss function.

    Parameters
    ----------
    yolo_output : tensor
        Final convolutional layer features.

    true_boxes : tensor
        Ground truth boxes tensor with shape [batch, num_true_boxes, 5]
        containing box x_center, y_center, width, height, and class.

    detectors_mask : array
        0/1 mask for detector positions where there is a matching ground truth.

    matching_true_boxes : array
        Corresponding ground truth boxes for positive detector positions.
        Already adjusted for conv height and width.

    anchors : tensor
        Anchor boxes for model.

    num_classes : int
        Number of object classes.

    rescore_confidence : bool, default=False
        If true then set confidence target to IOU of best predicted box with
        the closest matching ground truth box.

    print_loss : bool, default=False
        If True then use a tf.Print() to print the loss components.

    Returns
    -------
    mean_loss : float
        mean localization loss across minibatch
    """
    (yolo_output, true_boxes, detectors_mask, matching_true_boxes) = args
    num_anchors = len(anchors)
    object_scale = 5
    no_object_scale = 1
    class_scale = 1
    coordinates_scale = 1
    pred_xy, pred_wh, pred_confidence, pred_class_prob = yolo_head(
        yolo_output, anchors, num_classes)

    # Unadjusted box predictions for loss.
    # TODO: Remove extra computation shared with yolo_head.
    yolo_output_shape = K.shape(yolo_output)
    feats = K.reshape(yolo_output, [
        -1, yolo_output_shape[1], yolo_output_shape[2], num_anchors,
        num_classes + 5
    ])
    pred_boxes = K.concatenate(
        (K.sigmoid(feats[..., 0:2]), feats[..., 2:4]), axis=-1)

    # TODO: Adjust predictions by image width/height for non-square images?
    # IOUs may be off due to different aspect ratio.

    # Expand pred x,y,w,h to allow comparison with ground truth.
    # batch, conv_height, conv_width, num_anchors, num_true_boxes, box_params
    pred_xy = K.expand_dims(pred_xy, 4)
    pred_wh = K.expand_dims(pred_wh, 4)

    pred_wh_half = pred_wh / 2.
    pred_mins = pred_xy - pred_wh_half
    pred_maxes = pred_xy + pred_wh_half

    true_boxes_shape = K.shape(true_boxes)

    # batch, conv_height, conv_width, num_anchors, num_true_boxes, box_params
    true_boxes = K.reshape(true_boxes, [
        true_boxes_shape[0], 1, 1, 1, true_boxes_shape[1], true_boxes_shape[2]
    ])
    true_xy = true_boxes[..., 0:2]
    true_wh = true_boxes[..., 2:4]

    # Find IOU of each predicted box with each ground truth box.
    true_wh_half = true_wh / 2.
    true_mins = true_xy - true_wh_half
    true_maxes = true_xy + true_wh_half

    intersect_mins = K.maximum(pred_mins, true_mins)
    intersect_maxes = K.minimum(pred_maxes, true_maxes)
    intersect_wh = K.maximum(intersect_maxes - intersect_mins, 0.)
    intersect_areas = intersect_wh[..., 0] * intersect_wh[..., 1]

    pred_areas = pred_wh[..., 0] * pred_wh[..., 1]
    true_areas = true_wh[..., 0] * true_wh[..., 1]

    union_areas = pred_areas + true_areas - intersect_areas
    iou_scores = intersect_areas / union_areas

    # Best IOUs for each location.
    best_ious = K.max(iou_scores, axis=4)  # Best IOU scores.
    best_ious = K.expand_dims(best_ious)

    # A detector has found an object if IOU > thresh for some true box.
    object_detections = K.cast(best_ious > 0.6, K.dtype(best_ious))

    # TODO: Darknet region training includes extra coordinate loss for early
    # training steps to encourage predictions to match anchor priors.

    # Determine confidence weights from object and no_object weights.
    # NOTE: YOLO does not use binary cross-entropy here.
    # 计算不存在目标时的损失
    no_object_weights = (no_object_scale * (1 - object_detections) *
                         (1 - detectors_mask))
    no_objects_loss = no_object_weights * K.square(-pred_confidence)

    if rescore_confidence:
        objects_loss = (object_scale * detectors_mask *
                        K.square(best_ious - pred_confidence))
    else:
        objects_loss = (object_scale * detectors_mask *
                        K.square(1 - pred_confidence))
    confidence_loss = objects_loss + no_objects_loss

    # 类别损失
    # Classification loss for matching detections.
    # NOTE: YOLO does not use categorical cross-entropy loss here.
    matching_classes = K.cast(matching_true_boxes[..., 4], 'int32')
    matching_classes = K.one_hot(matching_classes, num_classes)
    classification_loss = (class_scale * detectors_mask *
                           K.square(matching_classes - pred_class_prob))

    # 坐标损失
    # Coordinate loss for matching detection boxes.
    matching_boxes = matching_true_boxes[..., 0:4]
    coordinates_loss = (coordinates_scale * detectors_mask *
                        K.square(matching_boxes - pred_boxes))

    confidence_loss_sum = K.sum(confidence_loss)
    classification_loss_sum = K.sum(classification_loss)
    coordinates_loss_sum = K.sum(coordinates_loss)
    total_loss = 0.5 * (
        confidence_loss_sum + classification_loss_sum + coordinates_loss_sum)
    if print_loss:
        total_loss = tf.Print(
            total_loss, [
                total_loss, confidence_loss_sum, classification_loss_sum,
                coordinates_loss_sum
            ],
            message='yolo_loss, conf_loss, class_loss, box_coord_loss:')

    return total_loss

yolov2数据输出