search

Etienne-Meunier-Inria / EM-Flow-Segmentation

Implementation for paper : EM-driven unsupervised learning for efficient motion segmentation
GNU Affero General Public License v3.0
13 stars 1 forks source link

Code to Evaluate the Model on Multilabel data #4

Open Coronal-Halo opened 8 months ago

Coronal-Halo commented 8 months ago

Hi, I see that the released checkpoint is for binary motion segmentation only, however, your paper states that the results improve as you increase the number of motions (K). I wonder if there is a way I can run your code to verify this? How can I verify the improved results by increasing K?

Etienne-Meunier-Inria commented 8 months ago

Hi there!

Thank you for your interest in multi-motion segmentation!

I've just added the weights and evaluation for the model reported in the paper to the weight section of the repository (repo). You can evaluate these models as described in the repo, just make sure to replace the > 0.5 threshold with an argmax.

You can also train your own multimask model using the -L flag. For the evaluation you can refer to the module "evaluationmodule"

Let me know if you need any additional info.

Best, Etienne

Coronal-Halo commented 8 months ago

Thank you so much!