Guided Similarity Separation for Image Retrieval

Authors: Chundi Liu, Guangwei Yu, Maksims Volkovs, Cheng Chang, Himanshu Rai, Junwei Ma, Satya Krishna Gorti

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we demonstrate that our model is able to successfully learn a new descriptor space that significantly improves retrieval accuracy, while still allowing efficient inner product inference. Experiments on five public benchmarks show highly competitive performance with up to 24% relative improvement in m AP over leading baselines.
Researcher Affiliation Industry Chundi Liu Layer6 AI chundi@layer6.ai Guangwei Yu Layer6 AI guang@layer6.ai Cheng Chang Layer6 AI jason@layer6.ai Himanshu Rai Layer6 AI himanshu@layer6.ai Junwei Ma Layer6 AI jeremy@layer6.ai Satya Krishna Gorti Layer6 AI satya@layer6.ai Maksims Volkovs Layer6 AI maks@layer6.ai
Pseudocode No The paper describes procedures and equations but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Full code for this work is available here: https://github.com/layer6ai-labs/GSS.
Open Datasets Yes The popular Oxford [29] and Paris [30] datasets have recently been revised to include more difficult instances, correct annotation mistakes, and introduce new evaluation protocols [32]. The new datasets, referred to as ROxford and RParis, contain 4,993 and 6,322 database images respectively. There are 70 query images in each dataset, and depending on the complexity of the retrieval task evaluation is further partitioned into Easy, Medium and Hard tasks. ... We also evaluate our model on the INSTRE dataset [42] which is an instance-level image retrieval benchmark containing various objects such as buildings, toys and book covers in natural scenes.
Dataset Splits No The paper describes using public benchmarks like ROxford, RParis, and INSTRE, and states the model is trained in a 'fully unsupervised fashion'. It does not provide explicit training, validation, or test dataset splits for model learning or hyperparameter tuning within these datasets, as they are primarily used as retrieval benchmarks.
Hardware Specification Yes All experiments are conducted on a 20-core Intel(R) Xeon(R) CPU E5-2630 v4 @2.20GHz machine with NVIDIA V100 GPU.
Software Dependencies No The paper mentions using 'TensorFlow library [1]' and 'ADAM optimizer [22]' but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes After parameter sweeps, two-layer 2048 2048 2048 GCN architecture produced the best performance and is used in all experiments. To set the important β parameter we note that only image pairs with sufficiently high similarity scores should be pushed closer together in the GSS loss. This leads to a general procedure where we first compute distribution of the pairwise scores sij, then set β in the upper percentile of this distribution. We consistently found that using 98 th percentile worked well across all datasets. Figure 2a illustrates this procedure and shows pairwise score distributions for each of the three datasets together with selected β values. Other hyper-parameters are set as follows: α = 1, k = 5 for ROxford; α = 1, k = 5 for RParis; α = 1, k = 10 for INSTRE. All models are optimized using the ADAM optimizer [22] with default settings and weight initialization outlined in Section 3.2 with ϵ = 10 5.