Closed MandiZhao closed 7 months ago
Hi @MandiZhao, we have a task called "interaction field estimation" in the paper, which is a fully-supervised method to estimate hand-object distances/contacts.
You can find its training script here: https://github.com/zc-alexfan/arctic/blob/master/scripts_method/train.py where we prepared the targets for the model.
train.py
does not contain how target data was prepared.
I saw the dataset is constructed it's somewhere in src/datasets/arctic_dataset.py
, but I still can't find keys like target.dist.or
in ArcticDataset.getitem
anywhere, while these dist
keys are required in construct_mesh
from src/mesh_loaders/field.py
-- is it possible some part of the code did not get released?
@zc-alexfan unfortunately the links alone were not very helpful.
What I found after much digging around is that, I needed to get the global hand and object vertices per time-step, which can be found as from processed_data (processed_data['world_coord']['verts.right']
etc) after running process_seq.py
on raw_seqs
, and combine them with the hand/object mesh faces. Distance/contact field can then be extracted using knn on the vertices. Closing this issue for now.
Hi @MandiZhao , sorry I was on the bus when replying to the issue. Yes your overall approach is the same way that I used. This link shows how to use the knn to get the distance values for contact. Hope it clarifies.
How can I generate & export the per-step hand-object binary contact information? I see the scripts_method/visualize.py has an option to visualize it from model predictions, but I'm trying to generate GT contact from the raw data. Thanks in advance!