Image Clustering with External Guidance

Authors: Yunfan Li, Peng Hu, Dezhong Peng, Jiancheng Lv, Jianping Fan, Xi Peng

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that TAC achieves state-of-the-art performance on five widely used and three more challenging image clustering benchmarks, including the full Image Net-1K dataset.
Researcher Affiliation Collaboration 1School of Computer Science, Sichuan University, Chengdu, China 2AI Lab at Lenovo Research, Beijing, China.
Pseudocode No The paper describes the proposed method in prose and mathematical equations but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes The code can be accessed at https://github.com/ XLearning-SCU/2024-ICML-TAC.
Open Datasets Yes To evaluate the performance of our TAC, we first apply it to five widely-used image clustering datasets including STL10 (Coates et al., 2011), CIFAR-10 (Krizhevsky & Hinton, 2009), CIFAR-20 (Krizhevsky & Hinton, 2009), Image Net10 (Chang et al., 2017b), and Image Net-Dogs (Chang et al., 2017b). ... Thus, we further evaluate the proposed TAC on three more complex datasets with larger cluster numbers, including DTD (Cimpoi et al., 2014), UCF-101 (Soomro et al., 2012), and Image Net-1K (Deng et al., 2009).
Dataset Splits No Following recent deep clustering works (Van et al., 2020; Dang et al., 2021a), we train and evaluate TAC on the train and test splits, respectively. Table 1 also clearly shows only 'Training Split' and 'Test Split' columns for all datasets, without a separate validation split for hyperparameter tuning.
Hardware Specification Yes All experiments are conducted on a single Nvidia RTX 3090 GPU.
Software Dependencies No The paper mentions using a pre-trained CLIP model and Adam optimizer, but does not provide specific version numbers for software dependencies like Python, PyTorch, or other libraries.
Experiment Setup Yes We train f and g by the Adam optimizer with an initial learning rate of 1e 3 for 20 epochs, with a batch size of 512. We fix τ = 5e 3, ˆτ = 0.5, and α = 5.0 in all the experiments. The only exception is that on UCF-101 and Image Net-1K, we change ˆτ to 5.0, batch size to 8192, and training epochs to 100, catering to the large cluster number.