Deep Metric Learning for Multi-Label and Multi-Object Image Retrieval

Content-based image retrieval has been a hot topic among computer vision researchers for long time. There have many advances over the years, one of recent ones being deep metric learning, inspired by success neural networks in machine learning tasks. The goal is to extract good high-level features from pixel data using networks. These provide useful abstractions, which can enable algorithms perform visual comparison between images with human-like accuracy. To learn these features, supervised information similarity or relative often used. One important issue how define multi-label multi-object scenes images. Traditionally, pairwise defined based on presence single common label two However, this definition very coarse and not suitable data. Another mistake completely ignore multiplicity objects images, hence ignoring facet certain types datasets. In our work, we propose an approach representations both We introduce intuitive effective Jaccard coefficient, equivalent intersection union sets. Hence treat as continuous, opposed discrete quantity. incorporate into triplet loss adaptive margin, achieve mean average precision further show, recently proposed quantization method, that resulting feature be quantized whilst preserving similarity. also show performs better than previously cosine similarity-based metric. Our method outperforms several state-of-the-art methods benchmark