Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection
Authors: Gengyu Lyu, Bohang Sun, Xiang Deng, Songhe Feng
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results on various multi-label datasets demonstrate that our CLS is significantly superior to other state-of-the-art methods. Experiments Experiment Setup Datasets We employ three multi-label datasets, including MS-COCO (Lin et al. 2014), PASCAL VOC 2007 (Everingham et al. 2010) and NUS-WIDE (Chua et al. 2009). |
| Researcher Affiliation | Collaboration | Gengyu Lyu1,2,4, Bohang Sun1, Xiang Deng2,3, Songhe Feng*2,5 1 College of Computer Science, Beijing University of Technology 2 School of Computer Science and Technology, Beijing Jiaotong University 3 Department of Automation, Tsinghua University 4 Idealism Beijing Technology Co., Ltd. 5 Key Laboratory of Big Data & Artificial Intelligence in Transportation (Beijing Jiaotong University), Ministry of Education |
| Pseudocode | Yes | Algorithm 1 describes the training process of our proposed CLS, where we train two deep neural networks in a mini-batch manner. |
| Open Source Code | No | No explicit statement about releasing code or a link to a repository is provided in the paper. |
| Open Datasets | Yes | We employ three multi-label datasets, including MS-COCO (Lin et al. 2014), PASCAL VOC 2007 (Everingham et al. 2010) and NUS-WIDE (Chua et al. 2009). Since these datasets are fully annotated, we follow (Chen et al. 2021b) to randomly drop some positive labels to generate ML-PL datasets according to a dropping rate α, where α {25%, 50%, 75%} indicates the proportion of dropped positive labels. |
| Dataset Splits | Yes | We employ three multi-label datasets, including MS-COCO (Lin et al. 2014), PASCAL VOC 2007 (Everingham et al. 2010) and NUS-WIDE (Chua et al. 2009). Since these datasets are fully annotated, we follow (Chen et al. 2021b) to randomly drop some positive labels to generate ML-PL datasets according to a dropping rate α, where α {25%, 50%, 75%} indicates the proportion of dropped positive labels. Besides, we consider the extreme case of ML-PL, where each instance is annotated with only one relevant label. |
| Hardware Specification | Yes | We train our model in an end-to-end manner and accomplish all experiments on a computer with an Intel (R) Xeon (R) CPU E52620, 64 GB main memory, and two TITAN Xp GPUs. |
| Software Dependencies | No | The paper mentions that the method is implemented based on PyTorch, but does not provide a specific version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | For fair comparison, we adopt Resnet-50 (He et al. 2016) network pre-trained on Image Net (Deng et al. 2009) as feature extraction backbone for all methods. The input images are squished and randomly cropped into 224 224. Adam is used as the optimizer with a weight decay of 10 4. The batch size is set to 120 for all datasets. We run 100 epochs in total with an initial learning rate of 4 10 5 and decrease it to 0 using cosine decay. We adopt Binary Cross Entropy loss as our loss function. We set the ratio of selection rate Rh(e) and Rl(e) as follows: Rh(e) = 1 min{ e Ek τh, τh}, Rl(e) = 1 min{ e Ek τl, τl}, where τl = 0.02 for PASCAL VOC 2007 dataset, and τl = 0.06 for NUS-WIDE dataset and MS-COCO dataset. Ek is set to 10 for all datasets. The values of τh are {0.002, 0.003, 0.005, 0.005}, {0.005, 0.010, 0.012, 0.012}, and {0.006, 0.012, 0.020, 0.020} on PASCAL VOC 2007 dataset, MS-COCO dataset, and NUS-WIDE dataset under four configurations respectively, which are found through cross-validation. |