Deep Metric Learning

Advertising in digital media often requires recognition of critical scenes in videos for smart placement of brand advertisements. These critical scenes raise viewer anxiety and are a part of some parent activity. We distinguish them from the rest of non-critical scenes using an order-preserving fine-grained similarity metric that learns the required representations. The learned metric is tested in two novel tasks: video critical scene recognition and fine-grained video retrieval. To learn the metric, we proposed Pentuplet Loss and recently, an improved and more robust Radial Loss.

These losses exploit the concept of `Quadlet Sampling’ to mine data where each training sample is a tuple of query, positive, intermediate and negative samples. Finally, to ascertain the effectiveness of the loss in learning a deep metric for measuring similarities, we tested its performance against state-of-the-art baselines in the known tasks of fine-grained image retrieval and shot-boundary detection.