Discovering Multi-Label Actor-Action Association in a Weakly Supervised Setting

This repository provides a PyTorch implementation of the actor action assignment given label logits Discovering Multi-Label Actor-Action Association in a Weakly Supervised Setting.

Qualitative Results:

Prerequisite

• Python >= 3.6 or 3.7
• PyTorch >= 1.3
• Numpy >= 1.18.0
• Cplex >= 12.10
• pickle >= 4.0 (optional)

Tested with:

• Python == 3.6
• PyTorch == 1.4
• Numpy == 1.19.0
• Cplex == 12.10

Usage Details

from LabelSpaceAssignment import LabelSpaceAssignment
logits_a = torch.randn(5,7) ## No. of instances = 5 and No. of classes = 7
bag_label = torch.empty(1,7).random_(2) ## No. of bags = 1 and No. of classes = 7
labelSpace = LabelSpaceAssignment(7) ## No. of classes = 7
labels,scores,omega_indx = labelSpace.assignmentSingle(logits_a,bag_label)

Citation:

If you use the code, please cite

S. Biswas and J. Gall.
Discovering Multi-Label Actor-Action Association in a Weakly Supervised Setting.
In Asian Conference on Computer Vision (ACCV), 2020