How to train depth and pose together? about geonet HOT 6 CLOSED

1. It depends on your training splits/datasets. You should follow the data preprocessing guide, then specify the dataset_dir as the formatted data;
2. In the first stage (train_rigid mode), your saved model should include both depthnet & posenet. Thus you only need to specify one ckpt.

from geonet.

1. Does formatted data contain both datasets under directory and main.py understand which one is depth and pose?

from geonet.

I'm not sure what you mean. The dataset simply contains image sequences and intrinsic matrices. The depth and pose should be trained in a joint way on the same formatted dataset.

from geonet.

So, you are saying that I can choose depth or pose(single one is enough) for training at first stage? If not, how to deal with the train.txt and val.txt comes from both of them?

Then, train in train flow mode with only flow formatted data?

The best explanation would be giving abstract directory sample though, by giving random directory names.

from geonet.

As described in the paper, for different tasks we have different training/test splits, which is handled by our data preparation script data/prepare_train_data.py.

For instance, if we want to replicate the flow result, we should prepare the data according to kitti_raw_stereo split. Then all our experiments should be conducted on the same formatted dataset. As a result, the trained model should be evaluated only for the flow task.

You could read the experiment part of our paper more thoroughly for details.

from geonet.

Thanks a lot. I could not get the logic from paper, it is very clear now.

from geonet.