Nuisance-Label Supervision: Robustness Improvement by Free Labels
This work addresses robustness issues for models in domains like action recognition, but it is incremental as it builds on existing supervision techniques with free labels.
The paper tackles the problem of model sensitivity to nuisance factors irrelevant to the task, such as changes in clothes or background in activity recognition, by introducing a Nuisance-label Supervision (NLS) module that uses freely acquired labels from data augmentation and synthetic data to improve robustness, resulting in consistent improvements in robustness to image corruption and appearance change.
In this paper, we present a Nuisance-label Supervision (NLS) module, which can make models more robust to nuisance factor variations. Nuisance factors are those irrelevant to a task, and an ideal model should be invariant to them. For example, an activity recognition model should perform consistently regardless of the change of clothes and background. But our experiments show existing models are far from this capability. So we explicitly supervise a model with nuisance labels to make extracted features less dependent on nuisance factors. Although the values of nuisance factors are rarely annotated, we demonstrate that besides existing annotations, nuisance labels can be acquired freely from data augmentation and synthetic data. Experiments show consistent improvement in robustness towards image corruption and appearance change in action recognition.