Faster R-CNN源码阅读之五：Faster R-CNN/lib/rpn_msr/proposal_layer_tf.py

# -*- coding:utf-8 -*-
# --------------------------------------------------------
# Faster R-CNN
# Copyright (c) 2015 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Ross Girshick and Sean Bell
# --------------------------------------------------------

import numpy as np
import yaml
from fast_rcnn.config import cfg
from generate_anchors import generate_anchors
from fast_rcnn.bbox_transform import bbox_transform_inv, clip_boxes
from fast_rcnn.nms_wrapper import nms
import pdb

DEBUG = False
"""
Outputs object detection proposals by applying estimated bounding-box
transformations to a set of regular boxes (called "anchors").
"""


def proposal_layer(rpn_cls_prob_reshape, rpn_bbox_pred, im_info, cfg_key, _feat_stride=[16, ],
                   anchor_scales=[8, 16, 32]):
    # Algorithm:
    #
    # for each (H, W) location i
    #   generate A anchor boxes centered on cell i
    #   apply predicted bbox deltas at cell i to each of the A anchors
    # clip predicted boxes to image
    # remove predicted boxes with either height or width < threshold
    # sort all (proposal, score) pairs by score from highest to lowest
    # take top pre_nms_topN proposals before NMS
    # apply NMS with threshold 0.7 to remaining proposals
    # take after_nms_topN proposals after NMS
    # return the top proposals (-> RoIs top, scores top)
    # layer_params = yaml.load(self.param_str_)

    # 算法
    # 对(H, W)的每个位置i：
    #    在i的位置生成A个anchor boxes
    #    对位置i的每个anchor使用预测的bboxdeltas
    # 裁剪预测框
    # 移除宽度或者高度小于某一阈值的预测框
    # 按从最高到最低的分数对所有（提案，分数）对进行排序
    # 在NMS之前获取top pre_nms_topN个proposals
    # 对其余proposals应用阈值为0.7的NMS
    # NMS之后取出after_nms_topN个proposals
    # 取出最前面的proposals  (-> RoIs top, scores top)

    # 获取若干anchors，这些anchors并没有加上偏移量，shape：[n, 4](此处n为anchors的个数)
    _anchors = generate_anchors(scales=np.array(anchor_scales))

    # 获取前面生成anchors的个数
    _num_anchors = _anchors.shape[0]

    # [N, H, W, C]通道顺序更改为[N, C, H, W]
    rpn_cls_prob_reshape = np.transpose(rpn_cls_prob_reshape, [0, 3, 1, 2])
    rpn_bbox_pred = np.transpose(rpn_bbox_pred, [0, 3, 1, 2])
    # rpn_cls_prob_reshape = np.transpose(np.reshape(rpn_cls_prob_reshape,[1,rpn_cls_prob_reshape.shape[0],rpn_cls_prob_reshape.shape[1],rpn_cls_prob_reshape.shape[2]]),[0,3,2,1])
    # rpn_bbox_pred = np.transpose(rpn_bbox_pred,[0,3,2,1])

    # 获取第一张图片的尺寸等信息
    im_info = im_info[0]

    # batch size必须为1，因为每次只能feed一张图片进入网络
    assert rpn_cls_prob_reshape.shape[0] == 1, \
        'Only single item batches are supported'
    # cfg_key = str(self.phase) # either 'TRAIN' or 'TEST'
    # cfg_key = 'TEST'

    # 以下代码获取配置信息， 包括nms的信息
    pre_nms_topN = cfg[cfg_key].RPN_PRE_NMS_TOP_N
    post_nms_topN = cfg[cfg_key].RPN_POST_NMS_TOP_N
    nms_thresh = cfg[cfg_key].RPN_NMS_THRESH
    min_size = cfg[cfg_key].RPN_MIN_SIZE

    # the first set of _num_anchors channels are bg probs
    # the second set are the fg probs, which we want
    # 按照channel通道取出RPN预测的框属于前景的分数
    # 请注意，在18个channel中，前9个是框属于背景的概率，后9个才是属于前景的概率（即是一个物体目标的概率）
    scores = rpn_cls_prob_reshape[:, _num_anchors:, :, :]

    # 重新取个变量名，bbox_deltas代表了RPN网络输出的各个框的变换信息
    bbox_deltas = rpn_bbox_pred
    # im_info = bottom[2].data[0, :]

    # 调试时的信息输出，下同
    if DEBUG:
        print 'im_size: ({}, {})'.format(im_info[0], im_info[1])
        print 'scale: {}'.format(im_info[2])

    # 1. Generate proposals from bbox deltas and shifted anchors
    # 1. 从bbox deltas和平移之后的anchors上生成proposals

    # 由于之前将通道顺序改为了[N, C, H, W]，因此直接获取最后的两个值对应于height和width
    height, width = scores.shape[-2:]

    if DEBUG:
        print 'score map size: {}'.format(scores.shape)

    # Enumerate all shifts
    # 枚举所有的平移量
    # 产生所有的x的集合
    shift_x = np.arange(0, width) * _feat_stride
    # 产生所有的y的集合
    shift_y = np.arange(0, height) * _feat_stride
    # 生成网格
    shift_x, shift_y = np.meshgrid(shift_x, shift_y)
    # 产生所有偏移量
    shifts = np.vstack((shift_x.ravel(), shift_y.ravel(),
                        shift_x.ravel(), shift_y.ravel())).transpose()

    # Enumerate all shifted anchors:
    #
    # add A anchors (1, A, 4) to
    # cell K shifts (K, 1, 4) to get
    # shift anchors (K, A, 4)
    # reshape to (K*A, 4) shifted anchors
    # 枚举所有加上偏移量之后的anchors的坐标
    # 因为生成的anchors包含左上角和右下角的两个坐标（相对于坐标原点0）
    # 因此上述的所有偏移量第三列和第一列相同，第四列和第二列相同
    # 加上anchors之后就获得了所有的偏移之后的初始proposals
    A = _num_anchors
    K = shifts.shape[0]
    anchors = _anchors.reshape((1, A, 4)) + \
              shifts.reshape((1, K, 4)).transpose((1, 0, 2))
    anchors = anchors.reshape((K * A, 4))

    # Transpose and reshape predicted bbox transformations to get them
    # into the same order as the anchors:
    #
    # bbox deltas will be (1, 4 * A, H, W) format
    # transpose to (1, H, W, 4 * A)
    # reshape to (1 * H * W * A, 4) where rows are ordered by (h, w, a)
    # in slowest to fastest order
    #
    # （上面的a表示通道数目，也常常使用c表示通道数目）
    # 转置并重新整理（reshape）预测的bbox变换，以便将其和anchors保持统一顺序
    bbox_deltas = bbox_deltas.transpose((0, 2, 3, 1)).reshape((-1, 4))

    # Same story for the scores:
    #
    # scores are (1, A, H, W) format
    # transpose to (1, H, W, A)
    # reshape to (1 * H * W * A, 1) where rows are ordered by (h, w, a)
    # 和上面的bbox deltas同理
    scores = scores.transpose((0, 2, 3, 1)).reshape((-1, 1))

    # Convert anchors into proposals via bbox transformations
    # 在这里结合RPN的输出变换初始框的坐标，得到第一次变换坐标后的proposals
    proposals = bbox_transform_inv(anchors, bbox_deltas)

    # 2. clip predicted boxes to image
    # 2. 在这里将超出图像边界的proposals进行边界裁剪，使之在图像边界之内
    proposals = clip_boxes(proposals, im_info[:2])

    # 3. remove predicted boxes with either height or width < threshold
    # (NOTE: convert min_size to input image scale stored in im_info[2])
    # 3. 去除那些宽度或者高度小于一定阈值的预测框，并返回符合条件的预测框的索引
    # （注：将min_size转换为存储在im_info [2]中的输入图像比例）
    keep = _filter_boxes(proposals, min_size * im_info[2])

    # 保留符合条件的proposals和scores
    proposals = proposals[keep, :]
    scores = scores[keep]

    # 4. sort all (proposal, score) pairs by score from highest to lowest
    # 4. 按照score从大到小的顺序给（proposals，score）对进行排序
    # 5. take top pre_nms_topN (e.g. 6000)
    # 5. 取出score最高的pre_nms_topN个（proposals，score）对
    order = scores.ravel().argsort()[::-1]
    if pre_nms_topN > 0:
        order = order[:pre_nms_topN]
    proposals = proposals[order, :]
    scores = scores[order]

    # 6. apply nms (e.g. threshold = 0.7)
    # 6. 对（proposals，score）对应用nms
    # 7. take after_nms_topN (e.g. 300)
    # 7. 取出后after_nms_topN个（proposals，score）对
    # 8. return the top proposals (-> RoIs top)
    # 8. 返回最上面的N个proposals
    keep = nms(np.hstack((proposals, scores)), nms_thresh)
    if post_nms_topN > 0:
        keep = keep[:post_nms_topN]
    proposals = proposals[keep, :]
    scores = scores[keep]

    # Output rois blob
    # Our RPN implementation only supports a single input image, so all
    # batch inds are 0
    # 返回rois blob。由于batch size为1，因此所有的batch inds均为0

    # batch inds，全为0的仅为1列的numpy数组
    batch_inds = np.zeros((proposals.shape[0], 1), dtype=np.float32)

    # 横向连接batch inds和proposals
    blob = np.hstack((batch_inds, proposals.astype(np.float32, copy=False)))

    # 返回
    return blob
    # top[0].reshape(*(blob.shape))
    # top[0].data[...] = blob

    # [Optional] output scores blob
    # if len(top) > 1:
    #    top[1].reshape(*(scores.shape))
    #    top[1].data[...] = scores


def _filter_boxes(boxes, min_size):
    """Remove all boxes with any side smaller than min_size."""
    # 去除高度或者宽度小于某一特定值的boxes

    # 计算高度和宽度
    ws = boxes[:, 2] - boxes[:, 0] + 1
    hs = boxes[:, 3] - boxes[:, 1] + 1

    # 计算高度和宽度均高于某一值的boxes的索引
    keep = np.where((ws >= min_size) & (hs >= min_size))[0]

    # 返回索引数组
    return keep