Problem by training on own dataset

[Jaisme]

Thanks a lot for the nice work! I am quite new to this field of depth estimation and would like to train this model on our own dataset (obtained from small space with about 0-20cm depth).

What I have done til now:

1. Generated my own dataset class
2. Started a test training with following parameters (without datapath etc.)

   --batch_size
   2
   --num_workers
   8
   --width
   960
   --height
   512
   --max_depth
   20.0
   --num_epochs
   5
   --scheduler_step_size
   8
   --learning_rate
   0.0001
   --thre
   0.95
   --num_layers
   18

   I adapted some code for the training, such as the line 41 self.num_pose_frames = 2 # self.num_input_frames in the execute_func.py, since the pose network is implemented only for two pose frames so far.

However, unfortunately I got another error AttributeError: 'Trainer' object has no attribute 'matching_ids' in the function predict_poses (about line 897) also in execute_func.py. Does anyone know how to resolve this error?

Furthermore, I have some questions about training:

1. is it possible to train this model with monocular images like ManyDepth?
2. which hyperparameters do I need to adjust for our own dataset?

I really appreciate for any help. Please let me know if you need more information :)

[choyingw]

The matching_ids issue is a collision feature for test-time single/multiframe mode. I will try to fix it next week. The easy way should be comment out https://github.com/facebookresearch/DistDepth/blob/8e45d570ae4771692c57baeb2c7cb5ee24f90b6b/execute_func.py#L887 to https://github.com/facebookresearch/DistDepth/blob/8e45d570ae4771692c57baeb2c7cb5ee24f90b6b/execute_func.py#L918 and see if that works.

(1) Yes, in paper, we also show training on ManyDepth. The easiest way is to build on ManyDepth's code and add distillation loss, which is what We've done in the paper. We'll consider releasing the part of the training code later.

(2) the input resolution might affect the results; larger resolution might give more high-frequency details but more local receptive fields. The num_layers=18 might be too small to learn a good depth map

[choyingw]

I have updated more instructions for how to train on your own dataset. Please check the README if interested!

[Jaisme]

Thanks for your reply! I'll definitely try it :D