[MOT] add MOT data #2789
[MOT] add MOT data #2789
Conversation
| 'DecodeVideo', 'Resize_LetterBox', 'AugmentHSV', 'NormalizedBbox2PixelBbox', | ||
| 'RandomAffine', 'BboxXYWH2XYXY', | ||
| 'Gt2JDETargetThres', 'Gt2JDETargetMax', | ||
| ] |
There was a problem hiding this comment.
这里面感觉DecodeVideo，AugmentHSV，NormalizedBbox2PixelBbox，RandomAffine，BboxXYWH2XYXY都属于通用的预处理操作？可以放到operators里？
There was a problem hiding this comment.
这里面感觉DecodeVideo，AugmentHSV，NormalizedBbox2PixelBbox，RandomAffine，BboxXYWH2XYXY都属于通用的预处理操作？可以放到operators里？
原本是放在operators.py里的，那只有'Gt2JDETargetThres', 'Gt2JDETargetMax'需要分出去到mot_operators.py吗？，还是也放在batch_operators.py ？
There was a problem hiding this comment.
Done, mot_operators.py里只保留了'Resize_LetterBox', 'Gt2JDETargetThres', 'Gt2JDETargetMax'
|
|
||
|
|
||
| @register_op | ||
| class Norm2PixelBbox(BaseOperator): |
There was a problem hiding this comment.
Norm2PixelBbox, BboxCXCYWH2XYXY能否去除？
There was a problem hiding this comment.
config里的EvalReader有用到，是eval检测结果时用到的
| dataset_dir (str): root directory for dataset. | ||
| image_dir (str): directory for images. | ||
| anno_path (str): annotation file path. | ||
| image_lists (str|list): mot data image lists, muiti-source mot dataset. |
There was a problem hiding this comment.
给个示例更好，看着下面有‘image_lists’， ‘labels_with_ids’硬规则
|
|
||
|
|
||
| @register_mot_op | ||
| class Resize_LetterBox(BaseOperator): |
There was a problem hiding this comment.
注意Python Code Style
Resize_LetterBox -> ResizeLetterBox
Letter这里是啥意思呢？
There was a problem hiding this comment.
letterbox是resize的一种，保持图片的长宽比，剩下的部分填补灰色，以免暴力resize造成图片失真。ultralytics/yolov3#232
现有operators.py的4个OP各自稍微改动也可以组成letterbox的效果，但是不完全对齐，与其改动4个OP不如新加一个OP
| super(DecodeVideo, self).__init__() | ||
|
|
||
| def apply(self, sample, context=None): | ||
| im = sample['image'] |
There was a problem hiding this comment.
这里的 sample['image']是个video吗？
Decode的含义是解码，VideoCapture才包含解码。下面cvtColor是解码之后的color space的变化。这里这个命名不适合。
另外，DecodeVideo这个函数有用吗？
There was a problem hiding this comment.
sample['image']是图片，视频是在dataloader里解码为图片的，所以送入DecodeVideo的已经是numpy的图片。DecodeVideo在config的TestMOTReader里，制作demo时或者infer直接读视频去过模型得到结果， 因为读图片文件夹的话，时间长点的视频抽出来的图片数太多太多，也占空间。
已暂时去除。
ppdet/data/source/mot.py
Outdated
| img_index = 0 | ||
| for data_name in self.image_lists: | ||
| with open( | ||
| os.path.join(dataset_dir, 'image_lists', data_name), |
There was a problem hiding this comment.
这里的image_lists应该对应anno_path?
There was a problem hiding this comment.
这个严格意义上不是anno，就只是个图片路径的list
| !MOTDataSet | ||
| dataset_dir: dataset/mot | ||
| image_lists: ['citypersons.val', 'caltech.val'] # for detection | ||
| # image_lists: ['caltech.10k.val', 'cuhksysu.val', 'prw.val'] # for reid |
There was a problem hiding this comment.
这个在什么场景用到，如果有模型依赖这个的话，可以另加一个mot_reid.yml的数据集配置
There was a problem hiding this comment.
这个在什么场景用到，如果有模型依赖这个的话，可以另加一个
mot_reid.yml的数据集配置
eval 模型的指标可选择3种模式。eval检测使用['citypersons.val', 'caltech.val']，eval reid 使用['caltech.10k.val', 'cuhksysu.val', 'prw.val']，这两种是EvalDataset的配置(MOTDataSet)，就和其他检测模型的流程一样。eval 跟踪的指标MOTA，具体eval哪个跟踪数据集由运行命令行参数控制，是EvalMOTDataset的配置(ImageFolder)，每个视频序列是按图片文件夹形式去读取的。
eval检测或reid并不常用，个人觉得没有必要分出去再写一个
| """ Resize the image numpy. | ||
| """ | ||
| im = sample['image'] | ||
| h, w = sample['im_shape'] |
| if isinstance(self.anno_path, list): | ||
| for path in self.anno_path: | ||
| self.dataset_dir = get_dataset_path(self.dataset_dir, path, | ||
| self.image_dir) |
There was a problem hiding this comment.
这里self.dataset_dir会被重写覆盖？
There was a problem hiding this comment.
不会覆盖，只是检查各个数据集目录是否存在。self.dataset_dir在configs/datasets/mot.yml里一般设置为dataset/mot
* add mot data * fix operators, source * fix data source transform * fix parse_dataset register_op * fix scale_factor, RandomAffine * add assert for check * fix ci
* add mot data * fix operators, source * fix data source transform * fix parse_dataset register_op * fix scale_factor, RandomAffine * add assert for check * fix ci
* add mot data * fix operators, source * fix data source transform * fix parse_dataset register_op * fix scale_factor, RandomAffine * add assert for check * fix ci
No description provided.