Deep Metric Learning with Hierarchical Triplet Loss

Abstract

We present a novel hierarchical triplet loss (HTL) capable of automatically collecting informative training samples (triplets) via a defined hierarchical tree that encodes global context information. This allows us to cope with the main limitation of random sampling in training a conventional triplet loss, which is a central issue for deep metric learning. Our main contributions are two-fold. (i) we construct a hierarchical class-level tree where neighboring classes are merged recursively. The hierarchical structure naturally captures the intrinsic data distribution over the whole dataset. (ii) we formulate the problem of triplet collection by introducing a new violate margin, which is computed dynamically based on the designed hierarchical tree. This allows it to automatically select meaningful hard samples with the guide of global context. It encourages the model to learn more discriminative features from visual similar classes, leading to faster convergence and better performance. Our method is evaluated on the tasks of image retrieval and face recognition, where it outperforms the standard triplet loss substantially by 1%18%. It achieves new state-of-the-art performance on a number of benchmarks, with much fewer learning iterations.

Metric Learning With HORDE: High-Order Regularizer for Deep Embeddings

“…First, we show in the upper part of Table 1 that HORDE significantly improves three popular baselines (contrastive loss, triplet loss and binomial deviance). These improvements allow us to claim state of the art results for single model methods on CUB with 58.3% R@1 (compared to 57.1% R@1 for HTL [5]) and second best for CARS.…”

SoftTriple Loss: Deep Metric Learning Without Triplet Sampling

“…Margin [12] takes 128 dimension of embeddings and uses ResNet50 [4] as the backbone. HTL [3] sets the dimension of embeddings to 512 and reports the state-of-the-art result on the backbone of Inception. With the large number of embeddings, it is obvious that all methods outperform existing DML methods with 64 embeddings in Table 1.…”

A Weakly Supervised Adaptive Triplet Loss for Deep Metric Learning

“…This also allows the metric learning process to generalize well on cross-domain datasets (table 1). (3) Inspired by [2], We propose an online anchor nearest neighbor sampling (OANNS) method which addresses the batch sampling issue for large scale training data. Our method out-performs state-of-art on both our collected fashion retrieval benchmark and significantly boosts the performance on a cross-domain benchmark even without further tuning.…”

“…The major difference between OANNS and [2] is that we save the extra computation cost of computing image vectors for all images. During batch training, we take the image vectors from the feed-forward output and update product vectors for the current batch of products.…”

“…In theory, we save 50% computation cost for the feed-forwarding process. However, the freshness of similarity information is comparable to [2] image vectors are only up-to-date in the first batch of each epoch in their training process, then become more and more out-dated until the end of the epoch.…”