A Strong and Robust Baseline for Text-Image Matching
This work provides incremental improvements for researchers and practitioners in multimodal AI by enhancing baseline methods for text-image matching.
The paper tackles the problem of text-image matching by proposing a new kNN-margin loss for training to handle hard negatives and noise, and using Inverted Softmax and Cross-modal Local Scaling during inference to address hubness, resulting in large-margin metric improvements.
We review the current schemes of text-image matching models and propose improvements for both training and inference. First, we empirically show limitations of two popular loss (sum and max-margin loss) widely used in training text-image embeddings and propose a trade-off: a kNN-margin loss which 1) utilizes information from hard negatives and 2) is robust to noise as all $K$-most hardest samples are taken into account, tolerating \emph{pseudo} negatives and outliers. Second, we advocate the use of Inverted Softmax (\textsc{Is}) and Cross-modal Local Scaling (\textsc{Csls}) during inference to mitigate the so-called hubness problem in high-dimensional embedding space, enhancing scores of all metrics by a large margin.