NVlabs
Use training script train2 for NDDS data
Hello
I need NDDS Data to work on the "new" training script with tensorboard visualisation. It seems not possible yet. Do you consider to improve in this way? I want to use NDDS data for training and NVISII data for testing. NDDS is pretty performant in creating big datasets for training and NVISII delivers good looking testing datasets.
I had to check myself to see what I did. This is what I did:
- The old script (
train.py) can now process both NDSS and NVISII data, but it doesn't have the nice tensorboard visualization. - The new script (
train2/train.py) can only process NVISII data.
The second point is @WilkoLu 's problem.
However, I think the differences are small, so it should be easy to adapt train2 to the NDSS format. Here is what's currently failing:
- "visibility" must be > 0, but it's always 0 in my NDSS data: https://github.com/NVlabs/Deep_Object_Pose/blob/538d12bf1aef17d3f094eaefe4f82d46d69b0575/scripts/train2/utils_dope.py#L227
I think the best solution here would be to fix the NDSS json data (search/replace), because NDSS only stores visible objects (but always sets visibility to "0"), whereas "visibility = 0" in NVISII means that the object actually is occluded/out of view, so it must be discarded.
- "projected_cuboid" must have 9 points (8 cuboid corners + centroid); the NDSS data has 8 points, and the field "projected_cuboid_centroid" is stored separately: https://github.com/NVlabs/Deep_Object_Pose/blob/538d12bf1aef17d3f094eaefe4f82d46d69b0575/scripts/train2/utils_dope.py#L228
Here we could check whether we have 8 or 9 points in the json file, and if it's 8, read "projected_cuboid_centroid" and append it, like I did in the old train.py script:
https://github.com/NVlabs/Deep_Object_Pose/blob/538d12bf1aef17d3f094eaefe4f82d46d69b0575/scripts/train.py#L294-L300
I think in the long run we should unify train.py and train2, there's no point in keeping both around.
