Difficulty-aware Balancing Margin Loss for Long-tailed Recognition
Authors: Minseok Son, Inyong Koo, Jinyoung Park, Changick Kim
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments validate our design choices and showcase the effectiveness and robustness of our method. We comprehensively evaluated our method on the long-tailed CIFAR, Image Net-LT, and i Naturalist2018 datasets, and demonstrated its effectiveness through extensive experiments. |
| Researcher Affiliation | Academia | Minseok Son*, Inyong Koo*, Jinyoung Park, Changick Kim Korea Advanced Institute of Science and Technology EMAIL |
| Pseudocode | No | The paper describes the proposed method using mathematical formulations and descriptive text, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/quotation2520/dbm ltr |
| Open Datasets | Yes | We conducted experiments on four benchmark long-tailed datasets. Long-tailed CIFAR-10 and CIFAR-100. Image Net-LT. i Naturalist2018. |
| Dataset Splits | Yes | We sampled long-tailed CIFAR datasets from the original CIFAR-10 and CIFAR-100 (Krizhevsky, Hinton et al. 2009) datasets with imbalance factors of 10, 50, and 100 using an exponential down-sampling profile outlined in (Cao et al. 2019; Cui et al. 2019). Evaluations were performed on the original balanced test sets. Image Net-LT (Liu et al. 2019c) is a longtailed version of Image Net-1K (Deng et al. 2009), sampled from a Pareto distribution with α = 6. It comprises 1,000 categories and 115.8K training images, with an imbalanced factor of ρ = 1280/5. The i Naturalist2018 dataset (Van Horn et al. 2018) is a large-scale real-world dataset that features a highly long-tailed distribution with an imbalance factor of ρ = 1000/2. It includes approximately 437K training images and 24.4K validation images gathered from 8,142 finegrained species classes in the wild. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used for the experiments. |
| Software Dependencies | No | The paper mentions using specific optimizers (SGD) and schedulers (cosine annealing scheduler), and data augmentation techniques (Cutout, Auto Augment), but it does not specify any software libraries or frameworks with version numbers (e.g., PyTorch 1.x, TensorFlow 2.x, Python 3.x). |
| Experiment Setup | Yes | The SGD optimizer with a momentum of 0.9 and weight decay of 2 10 4 was employed, along with a learning rate warm-up for the first five epochs and a cosine annealing scheduler for gradual decay. Data augmentation strategies included Cutout (De Vries and Taylor 2017) and Auto Augment (Cubuk et al. 2019). For BCL, we used an initial learning rate of 0.15 and a batch size of 256. For all other methods, we used an initial learning rate of 0.1 and a batch size of 64. Training was conducted for 200 epochs for most methods, except for NCL, which was trained for 400 epochs. In the case of DRW, class-balanced loss is introduced after 160 epochs. We used a scaling factor s = 32 for all our experiments, and tuned the hyperparameter for margin scaling K within the range 0.1 to 0.3, adjusting it based on datasets and baselines. For Image Net-LT, we utilized Res Net-50 and Res Ne Xt-50 (Xie et al. 2017) as backbones and trained them for 90 epochs. For i Naturalist2018, we employed Res Net-50 and trained for 100 epochs. In both benchmarks, we set the scaling factor s to 30 and the margin scaling hyperparameter K to 0.1. |