SpadeLiu
commented
1 year ago Hi,
The model used in this paper is not so robust, if there is a large domain gap between KITTI and RobotCar, the model can not generalize well. You can set the window size which defines the range to calculate disparity probability weighted average as 0 if large disparity changes are not supposed, but you should retrain the model from scratch, including pre-training on SeneFlow. It is not suggested to adjust parameters only for testing. On the other hand, the best solution is to train a model on your dataset RobotCar. You can also try some domain generalized models, for example, please refer to the other project in my github page.
刘必阳 @.***
------------------ Original ------------------ From: @.>; Send time: Thursday, Dec 29, 2022 12:55 PM @.>; @.***>; Subject: [SpadeLiu/Lac-GwcNet] test on Oxford RobotCar (Issue #15)
Hi, I have tested on the stereo images of Oxford RobotCar with pretrained model kitti2015.pth directly. However, the output disparity images have some white or black patches with extremely unsmooth grayscale values, and some parts of the building have insignificant differences in disparity values. Have you done any tests on the robotcar dataset? Is there any parameter adjustment that can optimize the above problem?
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>
Hi, I have tested on the stereo images of Oxford RobotCar with pretrained model kitti2015.pth directly. However, the output disparity images have some white or black patches with extremely unsmooth grayscale values, and some parts of the building have insignificant differences in disparity values. Have you done any tests on the robotcar dataset? Is there any parameter adjustment that can optimize the above problem?