Deep Metric Learning with Spherical Embedding
Authors: Dingyi Zhang, Yingming Li, Zhongfei Zhang
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on deep metric learning, face recognition, and contrastive self-supervised learning show that the SEC-based angular space learning strategy significantly improves the performance of the state-of-the-art. |
| Researcher Affiliation | Academia | Dingyi Zhang1, Yingming Li1 , Zhongfei Zhang2 1College of Information Science & Electronic Engineering, Zhejiang University, China 2Department of Computer Science, Binghamton University, USA |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described, nor does it state that code is available. |
| Open Datasets | Yes | (1) Deep metric learning task: we employ four fine-grained image clustering and retrieval benchmarks, including CUB200-2011 [28], Cars196 [29], SOP [5], and In-Shop [30]... (2) Face recognition task: CASIA-Web Face [32] is employed as the training set while the testing sets include LFW [33], Age DB30 [34], CFP-FP [35], and Mega Face Challenge 1 [36]... (3) Contrastive self-supervised learning task: we follow the framework and settings in SimCLR [19] and evaluate on CIFAR-10 and CIFAR-100 datasets [38]. |
| Dataset Splits | Yes | We Follow the protocol in [5, 30] to split the training and testing sets for them as in Table 1. For CUB200-2011 and Cars196, we do not use the bounding box annotations during training and testing. (Table 1 includes 'Train Test' sample counts). CASIA-Web Face [32] is employed as the training set while the testing sets include LFW [33], Age DB30 [34], CFP-FP [35], and Mega Face Challenge 1 [36]. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer [27]', 'SGD with momentum 0.9', 'BN-Inception [13]', 'ResNet50 [37]', 'SimCLR [19]', and 'NT-Xent' loss, but does not specify any version numbers for these software dependencies. |
| Experiment Setup | Yes | We set batch size to 120 and embedding size to 512 for all methods and datasets. We use Adam optimizer [27]. The compared methods are vanilla triplet loss (m = 1.0), semihard triplet loss (m = 0.2) [4], normalized N-pair loss (s = 25) [10, 18], and multi-similarity loss (ϵ = 0.1, λ = 0.5, α = 2, β = 40) [11]... We set batch size to 256 and embedding size to 512 for all methods. We use SGD with momentum 0.9. The hyper-parameter s is set to 64 while m for sphereface, cosface, and arcface are 3, 0.35, and 0.45, respectively... NT-Xent with temperature 0.5 is the loss and the batch size is 256. |