Closed martinlyra closed 1 year ago
https://github.com/wenbowen123/iros20-6d-pose-tracking/blob/fcf714d09bc9c3f711e2230b32dfb7f3b1884e86/datasets.py#LL149C14-L149C14
we are normalizing the rotations to the range [-1,1] by the self.rot_normalizer
which was set by https://github.com/wenbowen123/iros20-6d-pose-tracking/blob/fcf714d09bc9c3f711e2230b32dfb7f3b1884e86/dataset_info.yml#L13
The normalizer means at test time, how much do you think will the max rotation degree be between neighboring frames.
In your case, you can increase the normalizer value if you believe you can have very large rotation motion (also harder/longer to train) or generate the synthetic data pair with smaller rotation difference.
The normalizer value I am looking for is in datasets.py
and inside TrackDataset
's constructor; rot_normalizer=5*pi/180
, right?
where did you get rot_normalizer=5*pi/180
, can you paste the line?
On the end of this constructor.
When I checked out whether it was being set by the dataset_info.yml
, I couldn't find max_rotation
being used for the rot_normalizer
Hi, I just pushed a fix. The predict.py
are still hard-coded to reproduce the previous results. But when you train on your own data, the train.py
should now use the params from the dataset_info.yml
Awesome! Thanks for helping us with this.
Starting with some context, I and my partner, Marcus, are students working on our master thesis with topic on object detection applications in ROS on our own robot. We used the README to guide us, so we may get a successful application on our problem, which includes our own 3d-scanned models that we wish to detect and track the poses of.
When I try to run the train.py after generating about 1000 training and 200 validation data points and creating the data pairs (297 training & 200 validation), the training fails due to an assertion:
I would like to understand what this means, and what we can do to avoid getting this error. We have been able to generate smaller datasets (5/2 and 50/20) and train those datasets with success.