search

cleinc / bts

From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation
GNU General Public License v3.0
628 stars 177 forks source link

I cannot the reproduce nyu performance with densenet161 model #111

Open songya opened 3 years ago

songya commented 3 years ago

I used the default parameter settings using the file named "argument_train_nyu.txt (densenet161_bts setting). I trained three times but the d1 performance had never reached to the accuracy you published. According to your paper, it should achieve d1 = 0.885 but my results are 0.875, 0.877, and 0.878. The only difference is a multiprocessing option. With 'on-line eval' setting, the multiprocessing option cannot work. I used lr=1e-4, weight decay = 1e-2, adam_eps = 1e-3. Could you explain why I cannot reproduce your result?

freshjh commented 3 years ago

The performance reported in the paper is produced by a big batchsize, 16. You can refer to the section of implementation detail for more information.