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