AdaSample: Adaptive Sampling of Hard Positives for Descriptor Learning

TL;DR

AdaSample introduces an adaptive sampling method that focuses on selecting informative hard positives during training, significantly improving descriptor learning performance in computer vision tasks.

Contribution

The paper presents a novel adaptive online batch sampler that emphasizes hard positive mining, enhancing triplet loss training for local descriptor learning.

Findings

Significant performance improvements on standard benchmarks.

Consistent gains over strong baseline methods.

Effective positive sampling strategy enhances triplet loss training.

Abstract

Triplet loss has been widely employed in a wide range of computer vision tasks, including local descriptor learning. The effectiveness of the triplet loss heavily relies on the triplet selection, in which a common practice is to first sample intra-class patches (positives) from the dataset for batch construction and then mine in-batch negatives to form triplets. For high-informativeness triplet collection, researchers mostly focus on mining hard negatives in the second stage, while paying relatively less attention to constructing informative batches. To alleviate this issue, we propose AdaSample, an adaptive online batch sampler, in this paper. Specifically, hard positives are sampled based on their informativeness. In this way, we formulate a hardness-aware positive mining pipeline within a novel maximum loss minimization training protocol. The efficacy of the proposed method is evaluated on several standard benchmarks, where it demonstrates a significant and consistent performance gain on top of the existing strong baselines.