Learning Intra-Batch Connections for Deep Metric Learning

Authors: Jenny Denise Seidenschwarz, Ismail Elezi, Laura Leal-Taixé

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

Reproducibility Variable Result LLM Response
Research Type Experimental We achieve state-of-the-art results on clustering and image retrieval on the CUB-200-2011, Cars196, Stanford Online Products, and In-Shop Clothes datasets. To underline the effectiveness of our approach, we further present an extensive ablation study.
Researcher Affiliation Academia 1Department of Computer Science, Technical University of Munich, Munich, Germany. Correspondence to: Jenny Seidenschwarz <j.seidenschwarz@tum.de>.
Pseudocode No The paper describes its methodology using text and mathematical equations, and provides an overview figure, but does not include explicit pseudocode or a clearly labeled algorithm block.
Open Source Code Yes To facilitate further research, we make available the code and the models at https: //github.com/dvl-tum/intra_batch.
Open Datasets Yes We conduct experiments on 4 publicly available datasets using the conventional splitting protocols (Song et al., 2016): CUB-200-2011 (Wah et al., 2011)... Cars196 (Krause et al., 2013)... Stanford Online Products (SOP) (Song et al., 2016)... In-Shop Clothes (Liu et al., 2016).
Dataset Splits Yes CUB-200-2011 (Wah et al., 2011) ... For training, we use the first 100 classes and for testing the remaining classes. Cars196 (Krause et al., 2013) ... We use the first 98 classes for training and the remaining classes for testing. Stanford Online Products (SOP) (Song et al., 2016) ... We use 11, 318 classes for training and the remaining 11, 316 classes for testing. In-Shop Clothes (Liu et al., 2016) ... We use 3, 997 classes for training, while the test set, containing 3, 985 classes, is split into a query set and a gallery set.
Hardware Specification Yes All the training is done in a single Titan X GPU
Software Dependencies No The paper states: 'We implement our method in Py Torch (Paszke et al., 2017) library.' While PyTorch is named, a specific version number is not provided (e.g., PyTorch 1.x).
Experiment Setup Yes We use embedding dimension of sizes 512 for all our experiments and low temperature scaling for the softmax cross-entropy loss function... We resize the cropped image to 227 227, followed by applying a random horizontal flip. During test time, we resize the images to 256 256 and take a center crop of size 227 227. We train all networks for 70 epochs using RAdam optimizer (Liu et al., 2020)... We use small mini-batches of size 50-100.