Deep Metric Learning with Self-Supervised Ranking

Authors: Zheren Fu, Yan Li, Zhendong Mao, Quan Wang, Yongdong Zhang1370-1378

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on three standard benchmarks show that our method significantly improves and outperforms the state-of-the-art methods on the performances of both retrieval and ranking by 2%-4%. Experiments Given a query, retrieval can be separated into two stages, first distinguishes positive neighbors, then ranks them according to the degree of similarities. We evaluate the performances of our method in terms of retrieving and ranking. Datasets We evaluate our proposed method on three widely-used datasets following the standard protocol (Oh Song et al. 2016).
Researcher Affiliation Collaboration 1University of Science and Technology of China, Hefei, China 2Kuaishou Technology, Beijing, China 3Beijing Research Institute, University of Science and Technology of China, Beijing, China
Pseudocode Yes Algorithm 1 Model training process with our method
Open Source Code No The paper does not provide any statements about releasing code or links to a code repository.
Open Datasets Yes (1) CUB-200-2011 (CUB) (Wah et al. 2011) contains 11,788 images of 200 species of birds. (2) Cars-196 (Cars) (Krause et al. 2013) contains 16,185 images of 196 car models. (3) Stanford Online Products (SOP) (Oh Song et al. 2016) contains 120,053 online product images of 22,634 categories sold on e Bay.com.
Dataset Splits No CUB-200-2011 (CUB) (Wah et al. 2011) ... We use 5,864 images of its first 100 classes for training and 5,924 images of the remaining classes for testing. Cars-196 (Cars) (Krause et al. 2013) ... We use 8,054 images of its first 98 classes for training and 8,131 images of the other classes for testing. Stanford Online Products (SOP) (Oh Song et al. 2016) ... We use 59,551 images of 11,318 classes for training and 60,502 images of the rest for testing.
Hardware Specification Yes We use Py Torch (Paszke et al. 2019) to implement our method on a single GTX 1080Ti GPU with 11GB memory.
Software Dependencies No We use Py Torch (Paszke et al. 2019) to implement our method on a single GTX 1080Ti GPU with 11GB memory.
Experiment Setup Yes We use Adam W (Loshchilov et al. 2017) optimizer with 4e 4 weight decay and 120 batch size. The initial learning rate is 10 4 and scaled up 10 times on output layers for faster convergence. Mini-batches are constructed with the balanced sampler. The hyper-parameters setting is: α = 0.05, β = 0.5, s = 12, λ = 1.0, γ = 0.8, ptask = 0.8, M = 20, N = 4.