lisa traffic sign 数据集训练

1：lisa下载地址：http://cvrr.ucsd.edu/LISA/datasets.html
2：解压缩后，使用python tools/splitAnnotationFiles.py将数据集划分成训练集和测试集的csv描述文件。
3：转化csv为json格式。

import mmcv
import os
import sys
import json
import cv2
import numpy as np
 
class_names = ['stop', 'speedLimitUrdbl', 'speedLimit25', 'pedestrianCrossing', 'speedLimit35', 'turnLeft', 'slow', 'speedLimit15', 'speedLimit45', 'rightLaneMustTurn', 'signalAhead', 'keepRight', 'laneEnds', 'school', 'merge', 'addedLane', 'rampSpeedAdvisory40', 'rampSpeedAdvisory45', 'curveRight', 'speedLimit65', 'truckSpeedLimit55', 'thruMergeLeft', 'speedLimit30', 'stopAhead', 'yield', 'thruMergeRight', 'dip', 'schoolSpeedLimit25', 'thruTrafficMergeLeft', 'noRightTurn', 'rampSpeedAdvisory35', 'curveLeft', 'rampSpeedAdvisory20', 'noLeftTurn', 'zoneAhead25', 'zoneAhead45', 'doNotEnter', 'yieldAhead', 'roundabout', 'turnRight', 'speedLimit50', 'rampSpeedAdvisoryUrdbl', 'rampSpeedAdvisory50', 'speedLimit40', 'speedLimit55', 'doNotPass', 'intersection']
  
def _2coco(files):
    coco_anno = {'info': {}, 'images': [], 'licenses': [], 'annotations': [], 'categories': []}
    coco_anno['categories'] = [{'supercategory': j, 'id': i+1, 'name': j} for i,j in enumerate(class_names)]
    img_id = 0
    anno_id = 0
    img_ids = mmcv.list_from_file(files)
    for i in range(1,len(img_ids)):
        im_i = img_ids[i].split(';')
        file_name = im_i[0]
        cf = 1
        for z in range(0,len(coco_anno['images'])):
            if file_name == coco_anno['images'][z]['file_name']:
                im_id = coco_anno['images'][z]['id']
                cf = 2
        if cf == 1:
            img_id += 1
            img_ = cv2.imread(file_name)
            ih,iw = img_.shape[:2]
            img_info = {}
            img_info['id'] = img_id
            img_info['file_name'] = file_name
            img_info['height'] = ih
            img_info['width'] = iw
            coco_anno['images'].append(img_info)
        cls_name = im_i[1]
        x1 = float(im_i[2])
        y1 = float(im_i[3])
        x2 = float(im_i[4])
        y2 = float(im_i[5])
        if x2 < x1 or y2 < y1:
            print('bbox not valid: ',file_name)
            continue
        anno_id += 1
        bb = [x1, y1, x2 - x1, y2 - y1]
        categery_id = class_names.index(cls_name) + 1
        area = (x2 - x1) * (y2 - y1)
        anno_info = {}
        anno_info['segmentation'] = []
        anno_info['area'] = area
        if cf == 2:
            anno_info['image_id'] = im_id
        if cf == 1:
            anno_info['image_id'] = img_id
        anno_info['bbox'] = bb
        anno_info['iscrowd'] = 0
        anno_info['category_id'] = categery_id
        anno_info['id'] = anno_id
        coco_anno['annotations'].append(anno_info)

    return coco_anno


_coco = _2coco('split1.csv')
data_dir = sys.path[0]
with open(os.path.join(data_dir, 'coco_train.json'), 'w') as f:
    json.dump(_coco, f, ensure_ascii=False, indent=2)
print('done')
_coco = _2coco('split2.csv')
with open(os.path.join(data_dir, 'coco_test.json'), 'w') as f:
    json.dump(_coco, f, ensure_ascii=False, indent=2)
print('done')

4：安装mmdetection。此部分安装mmcv比较麻烦，原因是之前安装的mmcv会和mmcv-full冲突，直接用conda建一个新环境吧。我弄了俩三小时才弄明白。。。
5：mmdetction，我在用的过程中，发现准确率只有在第一个epoch的第一次是百分之九十，其余全是100，后来看代码发现是rpn_head里的Num_class是1，我修改成47后，直接报吐核错误。把per_gpu降到1还不行，试了一下voc，发现num_class是40。其中有个Bug：

```python
    def load_annotations(self, ann_file):
        """Load annotation from XML style ann_file.

        Args:
            ann_file (str): Path of XML file.

        Returns:
            list[dict]: Annotation info from XML file.
        """

        data_infos = []
        print("ann_file:",ann_file)
        img_ids = mmcv.list_from_file(ann_file)
        for img_id in img_ids:
            # print(img_id)
            img_id = img_id.split(' ')[0]
            filename = f'JPEGImages/{img_id}.jpg'
            xml_path = osp.join(self.img_prefix, 'Annotations',
                                f'{img_id}.xml')
            tree = ET.parse(xml_path)

img_id = img_id.split(’ ')[0]，这行需要加上去，在xml_style.py文件中，不然会连 -1也作为图片名字读取了。
6：faster_rcnn pytorch 链接：https://github.com/jwyang/faster-rcnn.pytorch
需要注意：使用torch0.4+torchvision0.2.0，以及用cuda9.0，再报错重新make一下。