megvii-research / RealFlow

The official implementation of the ECCV 2022 Oral paper: RealFlow: EM-based Realistic Optical Flow Dataset Generation from Videos
Other
93 stars 11 forks source link

重复训练的精度问题 #12

Open aaaaalun opened 1 year ago

aaaaalun commented 1 year ago

您好,我目前的流程如下

  1. 用现有的模型直接合成我自己的数据(自己的视频生成光流)
  2. 然后使用上面生成的光流来训练我的RAFT模型
  3. 使用新的RAFT模型来合成自己的数据,然后继续训练
  4. 不断重复 这整个过程虽然自动化了很多,但是如果生成自己的数据的时候光流有问题,那不是将模型训练的越来越有问题? 这个问题该怎么去避免?
coolbeam commented 1 year ago

“这整个过程虽然自动化了很多,但是如果生成自己的数据的时候光流有问题,那不是将模型训练的越来越有问题?”

Reply: Hello, While there is a possibility of encountering issues when training the model with self-generated data and potentially flawed optical flow, it is not a big prominent concern when utilizing a large amount of video data. To mitigate this, implementing data filtering methods can help avoid such problems.