facebookresearch/fairmotion

Motion graph dead end

Opened this issue · 6 comments

ameliacode commented

Hi I am currently making large dataset (approximately 10 min long) out of motion graph task.

When I check the Scadiver paper, it says the graph backtrace until it reaches to unvisited node. However, the code here gets out of the loop when it meets node with no outgoing edge. and the result generates less frames than the input..

I did also check the paper, motion vae with dynamic controller. It says you have generated 10 min length with CMU dataset, which I had experimented earlier and it resulted in 1min length..

Is there anyway to generate large motion dataset with this??

Jungdam commented

One possibility is to consider pruning option (either wcc or scc). If MG is constructed with wcc (weakly connected component) or no pruning, it is very likely that you eventually meet the dead end. You could also consider scc (strongly connected component) pruning, which guarantees that you never meet any dead end. However, in this case, many data could be removed due to stronger pruning constraint (It is very data-dependent how much data will survive in the end).

Another possibility is to reduce 'base_length' or 'stride_length' so that it can generate more connections but this would require more time and computational resources.

ameliacode commented

One more thing to ask: data doesn’t blend with other motion data…. Its node connect from the same mocap dats.. Any idea for this??

Jungdam commented

I did not understand what 'blend with other motion data' mean in your case. Can you elaborate more or give some examples?

ameliacode commented

motion idx shows identical numbers when I generate mocap data. (11 bvh data were used.. but..)
It doesn't transits to other node from other motion data..

Jungdam commented

Have you investigated that the constructed graph really include edges to other motion data? Depending on your dataset, it might require to adjust parameters for comparing two nodes (i.e. approximately two poses).

ameliacode commented

I haven't check the whole node, but everytime it starts randomly, it shows identical motion idx when traversing the nodes
Does "num_comparison" parameter has to do with this?? its default number is 3