Open horizon0408 opened 5 years ago
Hi. Sorry for confusing, because I lost my model weights which corresponds to the results on my paper when I was preparing the code. I trained it again, so there’s slightly difference. The version a/b/c correspond to full-1,2,3 exactly.
Hi, @lelimite4444 Last time, I tried to reproduce the results of full-2. First, I trained the monodepth without 2warp loss about 40 epochs. Then, I used this model as pre-trianed model and further trained this pre-trianed model with 2warp loss about 80epochs. I obtain the following results. abs_rel, sq_rel, rms, log_rms, d1_all, a1, a2, a3 0.0660, 0.8078, 4.240, 0.152, 8.901, 0.944, 0.978, 0.989 These results are slightly worse than your results in paper. Did I ignore any details? In stage of two, we only load the pre-trianed model as initialization, and the learning rate is similar to the stage one. In addition, I saw your pre-trained model using PWC-Net, which was trained with 40epochs. Does it mean the PWC-Net is first trained with 40 epochs (no 2warp) as initialization and then further trained 40 epochs with 2warp?
@horizon0408 @zmlshiwo Hi, may I ask you a question? Because I emailed the author and he didn't reply. When the author make the occlusion estimation for optical flow in the #128 of utils/utils.py, which corresponds to the formula (6) in the paper, the value of mag_sq is not consistent with the formula. Is there any typo here, or he did it for the convenience of calculation?
Hi! I noticed that the pretrained model in the drive folder is not available anymore. Can someone share it again?
Thanks!
Hi!
Appreciate your code and work. I have small questions with the pre-trained model you provided. Is the model monodepth_ver_a/b/c correspond to the full-1/2/3 in the Table 1 and Table 2 in your paper? I run your test and evaluation script with monodepth_ver_a on KITTI2015 and get the results like: abs_rel, sq_rel, rms, log_rms, d1_all, a1, a2, a3 0.0648, 0.8067, 4.186, 0.157, 8.623, 0.946, 0.979, 0.990 which is slightly different from Table 1. So I'm not sure if the model is the corresponding one or I missed some details for evaluation.