Long-Tail Cross Modal Hashing

Authors: Zijun Gao, Jun Wang, Guoxian Yu, Zhongmin Yan, Carlotta Domeniconi, Jinglin Zhang

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show the superior robustness and performance of Lt CMH to competitive CMH methods (Wang et al. 2020a; Li et al. 2018; Yu et al. 2021) on benchmark datasets, especially for long-tail ones. We adopt the typical mean average precision (MAP) as the evaluation metric, and report the average results and standard deviation in Table 2 (long-tailed), Table 3 (balanced) and Table 4 (single-modality).
Researcher Affiliation Academia Zijun Gao1,2, Jun Wang2,*, Guoxian Yu1,2, Zhongmin Yan1,2, Carlotta Domeniconi3, Jinglin Zhang4 1School of Software, Shandong University, Jinan, China 2SDU-NTU Joint Centre for AI Research, Shandong University, Jinan, China 3Department of Computer Science, George Mason University, Fairfax, VA, USA 4School of Control Science and Engineering, Shandong University, Jinan, China
Pseudocode No The paper states: 'We illustrate the whole framework of Lt CMH in Figure 1, and defer its algorithmic procedure into the Supplementary file.' This means pseudocode or algorithm blocks are not directly included in the main paper.
Open Source Code Yes The code of Lt CMH is shared at www.sdu-idea.cn/codes.php?name=Lt CMH.
Open Datasets Yes There is no off-the-shelf benchmark long-tail multi-modal dataset for experiments, so we pre-process two hand-crafted multi-modal datasets (Flickr25K (Huiskes and Lew 2008) and NUS-WIDE (Chua et al. 2009)), to make them fit long-tail settings.
Dataset Splits No The paper mentions 'Nbase' and 'Nquery' for datasets (Table 1), which typically refers to training/query sets in hashing, but it does not explicitly provide percentages or sample counts for training, validation, and test splits needed for full reproduction, nor does it explicitly mention a 'validation' split.
Hardware Specification No The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies Yes We implement Lt CMH in Python 3.7 with the Mind Spore deep learning framework.
Experiment Setup Yes SGD is used to optimize model parameters. Learning rate of image and text feature extraction is set as 1e-1.5, the learning rate of individuality-commonality AE is set as 1e-2. Other hyper-parameters are set as: batch size=128, α=0.05, β=0.05, γ=1, η=1, dx and dy is equal to hash code length k, the max epoch is 500. Parameter sensitivity is studied in the Supplementary file.