I try to use the inverse warping code in the JDACS for textureless area like planary wall.
However, the model collapse as the warped img is perfect while the depth is all blank. I am assuming the inverse warping and the bilinear sampling, spatial_transformer has the robustness to warp perfect img without the depth if the img contains too texureless area and little parallax (video sequence img). As a result, the learner can not learn because the perfectly warped image creates an illusion it has done good progress but in practice, it is not. Any suggestions or ideas why this is happening?
Hi, thanks for the nice project.
I try to use the inverse warping code in the JDACS for textureless area like planary wall.
However, the model collapse as the warped img is perfect while the depth is all blank. I am assuming the inverse warping and the bilinear sampling, spatial_transformer has the robustness to warp perfect img without the depth if the img contains too texureless area and little parallax (video sequence img). As a result, the learner can not learn because the perfectly warped image creates an illusion it has done good progress but in practice, it is not. Any suggestions or ideas why this is happening?
Any help is much appreciated. Thanks
ref img:![image](https://github.com/ToughStoneX/Self-Supervised-MVS/assets/47421732/75e9bbd0-0a35-465a-8478-e82a486dbc31)
seemingly perfect warped img:![image](https://github.com/ToughStoneX/Self-Supervised-MVS/assets/47421732/6300572b-777b-4114-bda5-51d22f10108e)
depth is not learnt and model collapose happens:![image](https://github.com/ToughStoneX/Self-Supervised-MVS/assets/47421732/e497c7ac-29b8-4a5a-a85b-d68a72eedc7f)