voc2007_extract_car_2_kitti

voc2007_extract_car_label

将上一步提取到的car转化为KITTI格式
1,第一步
获取所有图片的名，同时删除后缀。我这里用了比较笨的方法就是先把图片名保存，然后在gedit里面替换
保存图片名python代码：

import os
dirlist  = os.listdir("/home/mahnx0/data/voc_extract_car/train_img/")
fp = open("ImageID.txt","w")
for name in dirlist:
fp.write(name)
fp.write("\n")
fp.close()

gedit查找.jpg并替换成空

替换后如下：

开始转换，python代码如下：

# coding: utf-8

# In[24]:

import cv2, os, shutil
import xml.etree.ElementTree as ET


# In[2]:

ann_DIR = "/home/mahnx0/data/voc_extract_car/train_lable"
img_DIR = "/home/mahnx0/data/voc_extract_car/train_img"
trainval_file = "/home/mahnx0/data/voc_extract_car/train0.list"
test_file = "/home/mahnx0/data/voc_extract_car/val0.list"


# In[42]:

def readXmlAnno(im_fn, ann_DIR):
    anno_pn = os.path.join(ann_DIR, im_fn+'.xml')
    #print 'On annotation: {}'.format(anno_pn)
    tree = ET.parse(anno_pn)
    root = tree.getroot()

    p_anno = {}
    size = root.find('size')
    d_size = {"width": size.find('width').text, 
              "height": size.find('height').text,
              "depth": size.find('depth').text
             }        
    p_anno['size'] = d_size

    l_obj = []
    for obj in root.findall('object'):
        d_obj = {"name": obj.find('name').text, "truncated": '0.0', "difficult": '0.0', "occluded":'0.0',
                 "xmin": float(obj.find('bndbox').find('xmin').text),
                 "ymin": float(obj.find('bndbox').find('ymin').text),
                 "xmax": float(obj.find('bndbox').find('xmax').text),
                 "ymax": float(obj.find('bndbox').find('ymax').text),
                }
        l_obj.append(d_obj)

    p_anno['l_obj'] = l_obj

    if len(l_obj) > 0:       
        return p_anno
    else:
        return None    

def convertToKitti(p_anno):
    l_annos = []
    for obj in p_anno['l_obj']:
        k_anno = {}
        k_anno["type"] = obj["name"]
        k_anno["truncated"] = obj["truncated"]
        k_anno["occluded"] = obj["occluded"]
        k_anno["alpha"] = '0.0'
        k_anno["bbox"] = "{:.1f} {:.1f} {:.1f} {:.1f}".format(obj["xmin"],obj["ymin"],obj["xmax"],obj["ymax"])
        k_anno["dimensions"] = "{:.1f} {:.1f} {:.1f}".format(0,0,0)
        k_anno["location"] = "{:.1f} {:.1f} {:.1f}".format(0,0,0)
        k_anno["rotation_y"] = '0.0'
        l_annos.append(k_anno)

    return l_annos



# ### Create our own kitti dataset

# In[23]:

k_train_img_DIR = "/home/mahnx0/data/voc_extract_car_2_kitti/train_img"
k_train_lab_DIR = "/home/mahnx0/data/voc_extract_car_2_kitti/train_label"
k_val_img_DIR = "/home/mahnx0/data/voc_extract_car_2_kitti/val_img"
k_val_lab_DIR = "/home/mahnx0/data/voc_extract_car_2_kitti/val_label"


# In[47]:

# Processing train data
#'''
with open(trainval_file) as in_f:
    for im_fn in in_f:
        #print 'Processing img: {}'.format(im_fn)
        im_fn = im_fn.split('\n')[0].split('\r')[0]
        p_anno = readXmlAnno(im_fn, ann_DIR)
        if p_anno != None:
            l_annos = convertToKitti(p_anno)

            k_anno_file = os.path.join(k_train_lab_DIR,im_fn+".txt")
            with open(k_anno_file, 'w') as out_f:
                for k_anno in l_annos:
                    #print k_anno
                    to_write = ""
                    to_write += str(k_anno['type'])+' '
                    to_write += str(k_anno['truncated'])+' '
                    to_write += str(k_anno['occluded'])+' '
                    to_write += str(k_anno['alpha'])+' '
                    to_write += str(k_anno['bbox'])+' '
                    to_write += str(k_anno['dimensions'])+' '
                    to_write += str(k_anno['location'])+' '
                    to_write += str(k_anno['rotation_y'])
                    #print to_write
                    out_f.write(to_write+'\n')
                out_f.close()

            # copy im file
            from_file = os.path.join(img_DIR,im_fn+".jpg")
            to_file = os.path.join(k_train_img_DIR,im_fn+".jpg")
            shutil.copy2(from_file, to_file)
            #break
    in_f.close()
#'''

# Processing test data
with open(test_file) as in_f:
    for im_fn in in_f:
        #print 'Processing img: {}'.format(im_fn)
        im_fn = im_fn.split('\n')[0].split('\r')[0]
        p_anno = readXmlAnno(im_fn, ann_DIR)
        if p_anno != None:
            l_annos = convertToKitti(p_anno)

            k_anno_file = os.path.join(k_val_lab_DIR,im_fn+".txt")
            with open(k_anno_file, 'w') as out_f:
                for k_anno in l_annos:
                    #print k_anno
                    to_write = ""
                    to_write += str(k_anno['type'])+' '
                    to_write += str(k_anno['truncated'])+' '
                    to_write += str(k_anno['occluded'])+' '
                    to_write += str(k_anno['alpha'])+' '
                    to_write += str(k_anno['bbox'])+' '
                    to_write += str(k_anno['dimensions'])+' '
                    to_write += str(k_anno['location'])+' '
                    to_write += str(k_anno['rotation_y'])
                    #print to_write
                    out_f.write(to_write+'\n')
                out_f.close()
            # copy im file
            from_file = os.path.join(img_DIR,im_fn+".jpg")
            to_file = os.path.join(k_val_img_DIR,im_fn+".jpg")
            shutil.copy2(from_file, to_file)
            #break
    in_f.close()
# In[ ]:

OK，至此voc数据集转化KITTI工作全部完成
之前把coco,voc,kitti都转化成voc同时提取出了car,用yolov3训练发现效果好到爆炸，但是考虑到嵌入式平台TX2，速度无法达到实时，v3优化后1080p只有5fps, 用tensorrt又不好推理自定义层，所以当前准备用KITTI数据集，利用nvidia detectnet实现实时检测效果

下一步：转化DetracMvtData数据集到KITTI格式，这个数据我已经转化成voc了。至此已经有60000张不同场景的大小目标数据