AdaSample: Adaptive Sampling of Hard Positives for Descriptor Learning
arXiv:1911.12110
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.
References in corpus (5)
- Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
- PyTorch: An Imperative Style, High-Performance Deep Learning Library
- Instance Normalization: The Missing Ingredient for Fast Stylization
- GeoDesc: Learning Local Descriptors by Integrating Geometry Constraints
- Minimizing the Maximal Loss: How and Why?