search

haofeixu / aanet

[CVPR'20] AANet: Adaptive Aggregation Network for Efficient Stereo Matching
Apache License 2.0
524 stars 102 forks source link

Number of outputs #9

Closed waychin-weiqin closed 4 years ago

waychin-weiqin commented 4 years ago

Hi, in your paper you mentioned three different scales are used (1/3, 1/6, 1/12). In term of loss function, you first bilinearly upsample your outputs and use the ground truth labels as well as the pseudo round truth labels together with the outputs to compute the loss of each resolution. Balancing term, lambda is also included. But at the end of section 4.2, you mentioned lambda1,2,3 = 1.0, lambda4 = 2/3, lambda5 = 1/3.

I am wondering why are there 5 losses if you have only 3 outputs? Maybe I have misinterpreted something in your paper.

haofeixu commented 4 years ago

There are also two refinement modules to upsample the 1/3 resolution prediction to 1/2 and original resolutions. You can check the caption of Fig. 2 and Sec. 3.3 in our paper for more details.

waychin-weiqin commented 4 years ago

Thank you!