Xiaoyang Guo
Xiaoyang Guo
I am also trying to combine "tracking and mapping", but the performance is worse than the reported results...
No actually.
For some sequences such as fr1/desk2, there are actually several frames (9 out of 640 for fr1/desk2) not matched with ground-truth poses. This causes large evaluation errors.
I tried to set align_corners=True for grid_sample, and I got more points (22287197), but still less than 26727801.
Please use eval.sh
The 3D CNN consumes a lot of memory, it is hard to train with larger batch size. A possible solution is to reduce the number of channels or layers, but...
please make sure you place the images into correct directory. The directories are different for 12 and 15.
You can check the data format instructions on the submission page.
You can compute the disparity map using 960x540. Once you got the output disparity map, resize it back to 4608x3456. Then multiply a scale factor 4.8 (4608 / 960)
I recommend you to try recent stereo papers which focus on improving the generalization ability of stereo models, such as AdaStereo by Song et. al. Once you trained your stereo...