Neighborhood-Adaptive Structure Augmented Metric Learning

Authors: Pandeng Li, Yan Li, Hongtao Xie, Lei Zhang1367-1375

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on six standard benchmarks with two kinds of embeddings, i.e., binary embeddings and real-valued embeddings, show that our method significantly improves and outperforms the state-of-the-art methods.
Researcher Affiliation Collaboration 1School of Information Science and Technology, University of Science and Technology of China, Hefei, China 2Kuaishou Technology, Beijing, China
Pseudocode No The paper does not contain pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide any links or explicit statements about the availability of open-source code for the methodology described.
Open Datasets Yes Three datasets for binary embeddings. CIFAR10 (Krizhevsky, Hinton et al. 2009) consists of 60,000 images in 10 categories, which is randomly divided into training and testing sets, with 5,000 and 1,000 images. NUS-WIDE (Chua et al. 2009) is a large-scale dataset with 269,648 web images in 81 concepts. We select 21 largest concepts with 186,577 images, split 10,500 samples for training and 2,100 samples for testing. FLICKR25K (Huiskes and Lew 2008) has 25,000 images collected from Flickr in 24 concepts. We randomly select 2,000 images for testing and 5,000 images for training. Three datasets for real-valued embeddings. CUB-200-2011 (CUB) (Wah et al. 2011) has 11,788 images of 200 bird species. We use the first 100 species (5,864 images) for training and the rest 100 species (5,924 images) for testing. Cars-196 (Cars) (Krause et al. 2013) has 16,185 images of 196 cars. We split the first 98 cars (8,054 images) for training and the rest 100 cars (8,131 images) for testing. Stanford Online Products (SOP) (Oh Song et al. 2016) contains 120,053 images of 22,634 online products. We use the first 11,318 products (59,551 images) for training and 11,316 products (60,502 images) for testing.
Dataset Splits Yes CIFAR10 ... is randomly divided into training and testing sets, with 5,000 and 1,000 images. NUS-WIDE ... split 10,500 samples for training and 2,100 samples for testing. FLICKR25K ... randomly select 2,000 images for testing and 5,000 images for training. CUB-200-2011 ... We use the first 100 species (5,864 images) for training and the rest 100 species (5,924 images) for testing. Cars-196 ... We split the first 98 cars (8,054 images) for training and the rest 100 cars (8,131 images) for testing. Stanford Online Products (SOP) ... We use the first 11,318 products (59,551 images) for training and 11,316 products (60,502 images) for testing.
Hardware Specification No The paper mentions 'GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC' but does not provide specific details on the GPU models, CPUs, or other hardware components used.
Software Dependencies No The paper mentions the use of 'SGD algorithm', 'Adam optimizer', and 'Inception with batch normalization' but does not specify version numbers for any software dependencies or libraries.
Experiment Setup Yes Setting for binary embeddings. ...The hyper-parameters setting is: k = 8, β = 0.03 and λ = 1. Setting for Real-valued embeddings. ...We employ Adam optimizer with 1e 5 weight decay and fix the batch size as 128. For Proxynca++, we follow the settings in the original paper. For CUB, Cars and SOP, the balance factor λ are 4, 2 and 1, respectively. Hyper-parameters: t = 3 and α = 10. ...Inception with batch normalization (Ioffe and Szegedy 2015) is used as backbone and the input images are cropped to 224 224.