OpenVNA: A Framework for Analyzing the Behavior of Multimodal Language Understanding System under Noisy Scenarios
The OpenVNA contains one of the most comprehensive video noise injection toolkits, covering the most cases in real world applications. Furthermore, the API facilitates the implementation of noise-based data augmentation techniques, offering researchers a versatile toolkit to enhance their investigations.
For feature level noise injection, random feature drop (random erasure of feature sequences with zero-padding vectors) and structural feature drop (erasing of feature sequences with zero-padding vectors in succession) are provided. Corresponding code is provided in ./feature_noise.py and some examples are shown in test_feat_noise.py.
python test_feat_noise.pyFor raw video level noise injection, eight categories of perturbations are supported. Corresponding code is provided in ./real_noise.py and some examples are shown in test_real_noise.py.
python test_real_noise.pyThe OpenVNA framework integrates eight baselines on two human-centered video understanding tasks into the standardized evaluation process. The modular pipeline makes it very easy to integrate new models for a reliable comparison with existing baselines.
The OpenVNA framework extends prior quantitative evaluation criteria to raw video-level noise. It also furnishes an interactive platform for evaluating at the instance level, encompassing case analysis and error analysis.
