SiamRPN代码解读--proposal selection部

写在前面，首先吐槽一下，从siamrpn（pysot）的代码来看，作者的代码能力实在一般，（也可能是我更一般），代码整体不够clean，可读性也不高，并且和paper中的一些details出入也比较大。但出于工作需要，无奈还是要好好研读一下。

一.核心函数track

这一部分的前提是通过init函数进行crop_img和Bbox的初始化，即第一帧的跟踪对象的选定，而track函数是针对第一帧以后的帧进行操作。

part one

        w_z = self.size[0] + cfg.TRACK.CONTEXT_AMOUNT * np.sum(self.size)
        h_z = self.size[1] + cfg.TRACK.CONTEXT_AMOUNT * np.sum(self.size)
        s_z = np.sqrt(w_z * h_z)
        scale_z = cfg.TRACK.EXEMPLAR_SIZE / s_z
        s_x = s_z * (cfg.TRACK.INSTANCE_SIZE / cfg.TRACK.EXEMPLAR_SIZE)
        x_crop = self.get_subwindow(img, self.center_pos,
                                    cfg.TRACK.INSTANCE_SIZE,
                                    round(s_x), self.channel_average)
        outputs = self.model.track(x_crop)

这一部分是针对第一帧的对象，选取search的范围。

part two

score = self._convert_score(outputs['cls'])
pred_bbox = self._convert_bbox(outputs['loc'], self.anchors)

这一部分是将siamrpn model的输出进行后处理。outputs的type为dict。
outputs['cls']为（1x10x17x17）的tensor；
outputs["loc"]为(1x20x17x17）的tensor

---

def _convert_score(self, score):
        score = score.permute(1, 2, 3, 0).contiguous().view(2, -1).permute(1, 0)
        score = F.softmax(score, dim=1).data[:, 1].cpu().numpy()
        return score

_convert_score将原输入(1x10x17x17）变为（1445x2）后，经过softmax函数，再取第二列,因为第二列代表的是正样本的概率

---

def _convert_bbox(self, delta, anchor):
        delta = delta.permute(1, 2, 3, 0).contiguous().view(4, -1)
        delta = delta.data.cpu().numpy()

        delta[0, :] = delta[0, :] * anchor[:, 2] + anchor[:, 0]
        delta[1, :] = delta[1, :] * anchor[:, 3] + anchor[:, 1]
        delta[2, :] = np.exp(delta[2, :]) * anchor[:, 2]
        delta[3, :] = np.exp(delta[3, :]) * anchor[:, 3]
        return delta

_convert_bbox是对anchor的后处理，这里的anchor由generate_anchor生成,这里的score_size为17，注释写在代码中，这里又要吐槽一下，很明显，Sensetime公布的这份代码，目的就是开源，但也仅仅局限在开源了，并没有对代码做太多的美化，这让我不禁有点小小的disappointed,毕竟，Sensetime给了我那么多美好的回忆（跑题了）

def generate_anchor(self, score_size):
        anchors = Anchors(cfg.ANCHOR.STRIDE,
                          cfg.ANCHOR.RATIOS,
                          cfg.ANCHOR.SCALES)
        #cfg.ANCHOR.STRIDE=8
        #cfg.ANCHOR.RATIOS=[0.33,0.5,1,2,3]
        #cfg.ANCHOR.SCALES=[8]
        anchor = anchors.anchors
        x1, y1, x2, y2 = anchor[:, 0], anchor[:, 1], anchor[:, 2], anchor[:, 3]
        anchor = np.stack([(x1+x2)*0.5, (y1+y2)*0.5, x2-x1, y2-y1], 1)
        total_stride = anchors.stride
        anchor_num = anchor.shape[0]
        anchor = np.tile(anchor, score_size * score_size).reshape((-1, 4))
        ori = - (score_size // 2) * total_stride
        xx, yy = np.meshgrid([ori + total_stride * dx for dx in range(score_size)],
                             [ori + total_stride * dy for dy in range(score_size)])
        xx, yy = np.tile(xx.flatten(), (anchor_num, 1)).flatten(), \
            np.tile(yy.flatten(), (anchor_num, 1)).flatten()
        anchor[:, 0], anchor[:, 1] = xx.astype(np.float32), yy.astype(np.float32)
        return anchor

这里的anchor返回的是（1445,4）的矩阵。
表示[网格坐标y,网格坐标x，anchor纵坐标，anchor横坐标]

 def _convert_bbox(self, delta, anchor):
        delta = delta.permute(1, 2, 3, 0).contiguous().view(4, -1)
        delta = delta.data.cpu().numpy()

        delta[0, :] = delta[0, :] * anchor[:, 2] + anchor[:, 0]
        delta[1, :] = delta[1, :] * anchor[:, 3] + anchor[:, 1]
        delta[2, :] = np.exp(delta[2, :]) * anchor[:, 2]
        delta[3, :] = np.exp(delta[3, :]) * anchor[:, 3]
        return delta

这里实际上是对anchor大小，位置的更新，可以参考这些公式。
返回的delta也就是更新完的anchors

part three

这一部分主要是添加了一些惩罚项，惩罚项分为尺度惩罚项和空间位置上的惩罚项，尺度的惩罚项由下述部分完成。

---

def change(r):
            return np.maximum(r, 1. / r)

        def sz(w, h):
            pad = (w + h) * 0.5
            return np.sqrt((w + pad) * (h + pad))

        # scale penalty
        s_c = change(sz(pred_bbox[2, :], pred_bbox[3, :]) /
                     (sz(self.size[0]*scale_z, self.size[1]*scale_z)))
        #output_size:1445x1

        # aspect ratio penalty
        r_c = change((self.size[0]/self.size[1]) /
                     (pred_bbox[2, :]/pred_bbox[3, :]))
        penalty = np.exp(-(r_c * s_c - 1) * cfg.TRACK.PENALTY_K)
        pscore = penalty * score

但这一部分Code和Paper是有出入的。

---

这一部分是添加了余弦窗，来做一个空间位置上的滤波（认为下一帧的跟踪目标不会离这一帧太远）

pscore = pscore * (1 - cfg.TRACK.WINDOW_INFLUENCE) +\
                  self.window * cfg.TRACK.WINDOW_INFLUENCE
best_idx = np.argmax(pscore)

以上为Hanning窗的数学表示。久违了，SAS。

part four

在前三个part得到分数最高的anchor后，会做一些smooth的操作。

bbox = pred_bbox[:, best_idx] / scale_z
#尺度变换回去
lr = penalty[best_idx] * score[best_idx] * cfg.TRACK.LR

cx = bbox[0] + self.center_pos[0]
cy = bbox[1] + self.center_pos[1]

# smooth bbox
width = self.size[0] * (1 - lr) + bbox[2] * lr
height = self.size[1] * (1 - lr) + bbox[3] * lr
#这一部分可以看论文里的表述

代码这一部分没有采用NMS，而是直接取了max，这和paper里也不一样。

part five

这一部分将前面挑出的最大框返回。

 cx, cy, width, height = self._bbox_clip(cx, cy, width,height,img.shape[:2])
# udpate state
self.center_pos = np.array([cx, cy])
self.size = np.array([width, height])

bbox = [cx - width / 2,
        cy - height / 2,
        width,
        height]
best_score = score[best_idx]

return {
       'bbox': bbox,
       'best_score': best_score
        }

其中self._bbox_clip 解析如下，最后一个了，顶住。

def _bbox_clip(self, cx, cy, width, height, boundary):
        cx = max(0, min(cx, boundary[1]))
        cy = max(0, min(cy, boundary[0]))
        width = max(10, min(width, boundary[1]))
        height = max(10, min(height, boundary[0]))
        return cx, cy, width, height

就是一个简单的规避Bbox出界的函数。

---

附源代码

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import numpy as np
import torch.nn.functional as F

from pysot.core.config import cfg
from pysot.utils.anchor import Anchors
from pysot.tracker.base_tracker import SiameseTracker


class SiamRPNTracker(SiameseTracker):
    def __init__(self, model):
        super(SiamRPNTracker, self).__init__()
        self.score_size = (cfg.TRACK.INSTANCE_SIZE - cfg.TRACK.EXEMPLAR_SIZE) // \
            cfg.ANCHOR.STRIDE + 1 + cfg.TRACK.BASE_SIZE
        self.anchor_num = len(cfg.ANCHOR.RATIOS) * len(cfg.ANCHOR.SCALES)
        hanning = np.hanning(self.score_size)
        window = np.outer(hanning, hanning)
        self.window = np.tile(window.flatten(), self.anchor_num)
        self.anchors = self.generate_anchor(self.score_size)
        self.model = model
        self.model.eval()

    def generate_anchor(self, score_size):
        anchors = Anchors(cfg.ANCHOR.STRIDE,
                          cfg.ANCHOR.RATIOS,
                          cfg.ANCHOR.SCALES)
        anchor = anchors.anchors
        x1, y1, x2, y2 = anchor[:, 0], anchor[:, 1], anchor[:, 2], anchor[:, 3]
        anchor = np.stack([(x1+x2)*0.5, (y1+y2)*0.5, x2-x1, y2-y1], 1)
        total_stride = anchors.stride
        anchor_num = anchor.shape[0]
        anchor = np.tile(anchor, score_size * score_size).reshape((-1, 4))
        ori = - (score_size // 2) * total_stride
        xx, yy = np.meshgrid([ori + total_stride * dx for dx in range(score_size)],
                             [ori + total_stride * dy for dy in range(score_size)])
        xx, yy = np.tile(xx.flatten(), (anchor_num, 1)).flatten(), \
            np.tile(yy.flatten(), (anchor_num, 1)).flatten()
        anchor[:, 0], anchor[:, 1] = xx.astype(np.float32), yy.astype(np.float32)
        return anchor

    def _convert_bbox(self, delta, anchor):
        delta = delta.permute(1, 2, 3, 0).contiguous().view(4, -1)
        delta = delta.data.cpu().numpy()

        delta[0, :] = delta[0, :] * anchor[:, 2] + anchor[:, 0]
        delta[1, :] = delta[1, :] * anchor[:, 3] + anchor[:, 1]
        delta[2, :] = np.exp(delta[2, :]) * anchor[:, 2]
        delta[3, :] = np.exp(delta[3, :]) * anchor[:, 3]
        return delta

    def _convert_score(self, score):
        score = score.permute(1, 2, 3, 0).contiguous().view(2, -1).permute(1, 0)
        score = F.softmax(score, dim=1).data[:, 1].cpu().numpy()
        return score

    def _bbox_clip(self, cx, cy, width, height, boundary):
        cx = max(0, min(cx, boundary[1]))
        cy = max(0, min(cy, boundary[0]))
        width = max(10, min(width, boundary[1]))
        height = max(10, min(height, boundary[0]))
        return cx, cy, width, height

    def init(self, img, bbox):
        """
        args:
            img(np.ndarray): BGR image
            bbox: (x, y, w, h) bbox
        """
        self.center_pos = np.array([bbox[0]+(bbox[2]-1)/2,
                                    bbox[1]+(bbox[3]-1)/2])
        self.size = np.array([bbox[2], bbox[3]])

        # calculate z crop size
        w_z = self.size[0] + cfg.TRACK.CONTEXT_AMOUNT * np.sum(self.size)
        h_z = self.size[1] + cfg.TRACK.CONTEXT_AMOUNT * np.sum(self.size)
        s_z = round(np.sqrt(w_z * h_z))

        # calculate channle average
        self.channel_average = np.mean(img, axis=(0, 1))

        # get crop
        z_crop = self.get_subwindow(img, self.center_pos,
                                    cfg.TRACK.EXEMPLAR_SIZE,
                                    s_z, self.channel_average)
        self.model.template(z_crop)

    def track(self, img):
        """
        args:
            img(np.ndarray): BGR image
        return:
            bbox(list):[x, y, width, height]
        """
        w_z = self.size[0] + cfg.TRACK.CONTEXT_AMOUNT * np.sum(self.size)
        h_z = self.size[1] + cfg.TRACK.CONTEXT_AMOUNT * np.sum(self.size)
        s_z = np.sqrt(w_z * h_z)
        scale_z = cfg.TRACK.EXEMPLAR_SIZE / s_z
        s_x = s_z * (cfg.TRACK.INSTANCE_SIZE / cfg.TRACK.EXEMPLAR_SIZE)
        x_crop = self.get_subwindow(img, self.center_pos,
                                    cfg.TRACK.INSTANCE_SIZE,
                                    round(s_x), self.channel_average)

        outputs = self.model.track(x_crop)

        score = self._convert_score(outputs['cls'])
        pred_bbox = self._convert_bbox(outputs['loc'], self.anchors)

        def change(r):
            return np.maximum(r, 1. / r)

        def sz(w, h):
            pad = (w + h) * 0.5
            return np.sqrt((w + pad) * (h + pad))

        # scale penalty
        s_c = change(sz(pred_bbox[2, :], pred_bbox[3, :]) /
                     (sz(self.size[0]*scale_z, self.size[1]*scale_z)))

        # aspect ratio penalty
        r_c = change((self.size[0]/self.size[1]) /
                     (pred_bbox[2, :]/pred_bbox[3, :]))
        penalty = np.exp(-(r_c * s_c - 1) * cfg.TRACK.PENALTY_K)
        pscore = penalty * score

        # window penalty
        pscore = pscore * (1 - cfg.TRACK.WINDOW_INFLUENCE) + \
            self.window * cfg.TRACK.WINDOW_INFLUENCE
        best_idx = np.argmax(pscore)

        bbox = pred_bbox[:, best_idx] / scale_z
        lr = penalty[best_idx] * score[best_idx] * cfg.TRACK.LR

        cx = bbox[0] + self.center_pos[0]
        cy = bbox[1] + self.center_pos[1]

        # smooth bbox
        width = self.size[0] * (1 - lr) + bbox[2] * lr
        height = self.size[1] * (1 - lr) + bbox[3] * lr

        # clip boundary
        cx, cy, width, height = self._bbox_clip(cx, cy, width,
                                                height, img.shape[:2])

        # udpate state
        self.center_pos = np.array([cx, cy])
        self.size = np.array([width, height])

        bbox = [cx - width / 2,
                cy - height / 2,
                width,
                height]
        best_score = score[best_idx]
        return {
                'bbox': bbox,
                'best_score': best_score
               }