Fast hard negative mining for deep metric learning
作者：
Highlights：
• Bag of Negatives - a computationally inexpensive and mini-batch size independent method for improved negative mining in large datasets.
• Images used for training of Siamese networks are sampled from hash bins, which contain images that are close in embedding space.
• Hash table is dynamically updated in every training step with minor computational overhead.
• BoN speeds up training three times, while providing significantly better accuracy.
摘要
•Bag of Negatives - a computationally inexpensive and mini-batch size independent method for improved negative mining in large datasets.•Images used for training of Siamese networks are sampled from hash bins, which contain images that are close in embedding space.•Hash table is dynamically updated in every training step with minor computational overhead.•BoN speeds up training three times, while providing significantly better accuracy.
论文关键词：Deep metric learning,Instance retrieval,Re-identification,Siamese networks,Online hashing
论文评审过程：Received 5 November 2019, Revised 28 September 2020, Accepted 13 December 2020, Available online 17 December 2020, Version of Record 3 January 2021.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107795
