Probability Guided Loss for Long-Tailed Multi-Label Image Classification
Authors: Dekun Lin
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on two long-tailed multi-label image classification datasets: VOC-LT and COCO-LT. The results demonstrate the rationality and superiority of our strategy. |
| Researcher Affiliation | Academia | Chengdu Institute of Computer Applications, Chinese Academy of Sciences, Chengdu 610041, China University of Chinese Academy of Sciences, Beijing 100049, China |
| Pseudocode | No | The paper presents mathematical equations for its proposed loss functions but does not include any pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states: "All the codes of DB loss are took out and retrained by us." This refers to third-party code used, not code developed by the authors for their proposed method. No explicit statement or link to the authors' own source code is provided. |
| Open Datasets | Yes | We conduct experiments on two long-tailed multi-label classification datasets: VOC-LT and COCO-LT. They are artificially constructed from Pascal Visual Object Classes Challenge (VOC) (Everingham et al. 2015) and MS-COCO (Lin et al. 2014), respectively. |
| Dataset Splits | No | The paper specifies the training and testing set sizes for VOC-LT and COCO-LT but does not explicitly detail a separate validation split or how validation was handled (e.g., specific percentage, number of samples, or citation to a predefined validation split). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments, such as GPU models, CPU specifications, or memory. |
| Software Dependencies | Yes | We conduct all experiments on Py Torch1.8.0. |
| Experiment Setup | Yes | We use the Res Net50 (He et al. 2016) which is pre-trained on Image Net (Deng et al. 2009) as the backbone of model. The input images are all resized to a dimension of 224 224 and the batch size is 32. We adopt standard data augmentations the same as DB loss. The optimizer we take is SGD whose momentum is 0.9 and weightdecay is 1e-4. The initial lr is 8e-3 for VOC-LT and 1e-2 for COCO-LT. We also use warm-up learning rate schedule (Goyal et al. 2017) for the first 500 iterations with a ratio of 1/3. |