ALBAR trains a single video encoder with an adversarial objective on repeated-frame static clips, entropy maximization and a gradient penalty, raising HMDB51 SCUBA/SCUFO contrasted accuracy from 40.28% to 53.22% and introducing a leakage-free UCF101 protocol.
16 Published as a conference paper at ICLR 2025 Table 6: Results on IID and OOD test sets of various debiasing methods on Kinetics400
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
ALBAR: Adversarial Learning approach to mitigate Biases in Action Recognition
ALBAR trains a single video encoder with an adversarial objective on repeated-frame static clips, entropy maximization and a gradient penalty, raising HMDB51 SCUBA/SCUFO contrasted accuracy from 40.28% to 53.22% and introducing a leakage-free UCF101 protocol.