The field of computer vision is moving towards improving the robustness and reliability of deep neural networks, particularly in out-of-distribution (OOD) detection and vision-language alignment. Researchers are exploring innovative methods to identify and mitigate biases in convolutional neural networks (CNNs) and to enhance the generalization ability of vision-language models (VLMs) to covariate-shifted OOD data. Noteworthy papers in this area include:
- One study that proposes techniques for identifying hidden biases in CNNs, using image scrambling and transforms to distinguish between contextual information and background noise.
- Another study that introduces a novel OOD score, ΔEnergy, which significantly outperforms existing methods and provides a reliable approach for OOD detection and generalization.
- A method that utilizes local background features as fake OOD features for model training, achieving state-of-the-art performance in OOD detection benchmarks.