Faster RCNN 结果可视化

visible_faster_rcnn_result.py

# -*- coding: utf-8 -*-
import pandas as pd
import numpy as np
import os

from sklearn.externals import joblib
from skimage import io, draw, img_as_float

pkl_pred = "py-faster-rcnn/output/faster_rcnn_alt_opt/voc_2007_test/VGG16_faster_rcnn_final/detections.pkl"
pkl_anno = "py-faster-rcnn/data/VOCdevkit2007/annotations_cache/annots.pkl"
txt_test = "py-faster-rcnn/data/VOCdevkit2007/VOC2007/ImageSets/Main/test.txt"
img_dir = "py-faster-rcnn/data/VOCdevkit2007/VOC2007/JPEGImages"

# 重合超过多少就算正确
ovthresh = 0.3

# 总共检测到多少
total_predicted = 0

# 实际有多少
total_groundtruth = 0

# 对的
counter_yes = 0

# 错的
counter_no = 0

predictions = joblib.load(pkl_pred)[1]
annotations = joblib.load(pkl_anno)


def fnms(txt):
    with open(txt, 'r') as test_set:
        for ith, fnm in enumerate(test_set):
            yield (ith, fnm.strip())


def vis(img, bbgt, bb, yes, no):
    # 可视化
    image_data = img_as_float(io.imread(img))
    # 蓝色是GroundTruth
    for b in bbgt:
        rr, cc = draw.polygon_perimeter([int(b[1]), int(b[1]), int(b[3]), int(b[3])],
                                        [int(b[0]), int(b[2]), int(b[2]), int(b[0])])
        draw.set_color(image_data, (rr, cc), (0, 0, 1))

    for ith, b in enumerate(bb):
        rr, cc = draw.polygon_perimeter([int(b[1]), int(b[1]), int(b[3]), int(b[3])],
                                        [int(b[0]), int(b[2]), int(b[2]), int(b[0])])
        # 绿色是预测正确
        # 红色是预测错误
        if ith in yes:
            draw.set_color(image_data, (rr, cc), (0, 1, 0))
        else:
            draw.set_color(image_data, (rr, cc), (1, 0, 0))
    io.imshow(image_data)
    io.show()


# 对于每一张图片
for ith, fnm in fnms(txt_test):
    img = os.path.join(img_dir, fnm + '.png')
    # 获取到ground truth和预测结果
    gt = annotations[fnm]
    pr = predictions[ith]

    # 统计个数
    total_predicted += len(pr)
    total_groundtruth += len(gt)

    yes = list()
    yes_pair = list()
    no = list()

    # 做计算
    # 对于每一个检测出来的框子
    BBGT = np.array([x['bbox'] for x in gt])
    for ith, bbpr in enumerate(pr):
        bb = np.array(bbpr)
        ovmax = -np.inf
        if len(gt) > 0:
            ixmin = np.maximum(BBGT[:, 0], bb[0])
            iymin = np.maximum(BBGT[:, 1], bb[1])
            ixmax = np.minimum(BBGT[:, 2], bb[2])
            iymax = np.minimum(BBGT[:, 3], bb[3])
            iw = np.maximum(ixmax - ixmin + 1., 0.)
            ih = np.maximum(iymax - iymin + 1., 0.)
            inters = iw * ih

            # union
            uni = ((bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.) +
                   (BBGT[:, 2] - BBGT[:, 0] + 1.) *
                   (BBGT[:, 3] - BBGT[:, 1] + 1.) - inters)

            overlaps = inters / uni
            ovmax = np.max(overlaps)
            jmax = np.argmax(overlaps)
            yes_pair.append(jmax)

        if ovmax > ovthresh:
            counter_yes += 1
            yes.append(ith)
        else:
            counter_no += 1
            no.append(ith)
    print("Image:", fnm)
    print("Truth:", len(gt))
    print("Yes:", len(yes))
    print("No:", len(no))
    print("Miss", len(gt) - len(yes))
    print()
    vis(img, BBGT, pr, yes, no)

print()
print("total_detected", total_predicted)
print("total_groundtruth", total_groundtruth)
print("Counter_yes", counter_yes)
print("Counter_no", counter_no)
print("Counter_miss", total_groundtruth - counter_yes)

visible_loss.py

# -*- coding: utf-8 -*-

import re
import numpy as np
import matplotlib.pyplot as plt


def str2num(str_list, output_type):
    return [output_type(x) for x in str_list]


if "__main__" == __name__:
    log_file = "experiments/logs/faster_rcnn_end2end_VGG16_.txt.2017-04-11_10-32-10"
    pattern_itr = re.compile(r"219\]\s+Iteration\s+([\d]+)")
    pattern_loss = re.compile(r", loss[\s=]{1,3}([\d\.]+)")
    # pattern_loss = re.compile(r" loss_bbox[\s=]{1,3}([\d\.]+)")
    # pattern_loss = re.compile(r" loss_cls[\s=]{1,3}([\d\.]+)")
    # pattern_loss = re.compile(r" rpn_cls_loss[\s=]{1,3}([\d\.]+)")
    # pattern_loss = re.compile(r" rpn_loss_bbox[\s=]{1,3}([\d\.]+)")

    with open(log_file, 'r') as f:
        lines = f.read()
        itrs = pattern_itr.findall(lines)
        loss = pattern_loss.findall(lines)

        itrs = np.array(str2num(itrs, int))
        loss = np.array(str2num(loss, float))

        plt.figure()
        plt.plot(itrs, loss)
        plt.show()