search

Y0uchenZ / TANet

A Fast Stereo Matching Network Based on Temporal Attention and 2D Convolution
MIT License
8 stars 1 forks source link

disp_pre is ground truth? #1

Open hodakagoto opened 9 months ago

hodakagoto commented 9 months ago

Hi, is it using ground truth for disp_pre not the inferred output, both in training and in testing?

Y0uchenZ commented 9 months ago

Hi, the "disp_pre" is calculated by other network as additional data, not directly using gt.

Hi, is it using ground truth for disp_pre not the inferred output, both in training and in testing?

hodakagoto commented 9 months ago

Thank you. Sorry for not realizing that the README mentioned that:

In addition to the publicly available KITTI dataset, we used an additional dataset, namely the disparity maps of the previous frame (computed using PSMNet).

Is there any specific reason for using PSMNet for disp_pre? I think it's natural to use the previous output from the same model for disp_pre.

Y0uchenZ commented 9 months ago

In fact, we just randomly selected a classic network (such as PSMNet) to supplement the dataset. The structure of TANet requires the previous frame disparity map as input, but during the training process, we can only use the results of other algorithms for training (after all, it is not self supervised). In practical applications, the input of the current frame can definitely use the TANet's output of the previous frame.

Thank you. Sorry for not realizing that the README mentioned that:

In addition to the publicly available KITTI dataset, we used an additional dataset, namely the disparity maps of the previous frame (computed using PSMNet).

Is there any specific reason for using PSMNet for disp_pre? I think it's natural to use the previous output from the same model for disp_pre.