Multi-level Distance Regularization for Deep Metric Learning
Authors: Yonghyun Kim, Wonpyo Park1827-1835
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We extensively perform ablation studies on its behaviors to show the effectiveness of MDR. Without bells and whistles, MDR with simple Triplet loss achieves the-state-of-the-art performance in various benchmark datasets: CUB-200-2011, Cars-196, Stanford Online Products, and In-Shop Clothes Retrieval. |
| Researcher Affiliation | Industry | Yonghyun Kim1*, Wonpyo Park2* 1AI Lab, Kakao Enterprise 2Kakao Corp. aiden.d@kakaoenterprise.com, tony.nn@kakaocorp.com |
| Pseudocode | No | The paper describes the detailed procedure of MDR with equations and figures, but it does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code or provide a link to a code repository. |
| Open Datasets | Yes | We employ the four standard datasets of deep metric learning for evaluations: CUB-200-2011 (Wah et al. 2011) (CUB-200), Cars-196 (Krause et al. 2013), Stanford Online Product (Oh Song et al. 2016) (SOP) and In-Shop Clothes Retrieval (Liu et al. 2016) (In-Shop). |
| Dataset Splits | Yes | We follow the standard evaluation protocol and data splits proposed in (Oh Song et al. 2016). For an unbiased evaluation, we conduct 5 independent runs for each experiment and report the mean and the standard deviation of them. CUB-200 has 5,864 images of first 100 classes for training and 5,924 images of the rest classes for evaluation. Cars-196 has 8,054 images of first 98 classes for training and 8,131 images of the rest classes for evaluation. SOP has 59,551 images of 11,318 classes for training and 60,502 images of the rest classes for evaluation. In-Shop has 25,882 images of 3,997 classes for training, and the remaining 7,970 classes with 26,830 images are partitioned into two subsets (query set and gallery set) for evaluation. |
| Hardware Specification | No | The paper mentions "Inception architecture with Batch Normalization (IBN)" and "Res Net18 (R18) and Res Net50 (R50)" as backbone networks, but these are model architectures, not hardware specifications. There is no mention of specific GPUs, CPUs, or other hardware used for training or inference. |
| Software Dependencies | No | The paper mentions "Adam (Kingma and Ba 2014) optimizer" and "Triplet loss (Schroff, Kalenichenko, and Philbin 2015)", but it does not specify software versions for any libraries, frameworks (e.g., TensorFlow, PyTorch), or programming languages used for implementation. |
| Experiment Setup | Yes | We employ Adam (Kingma and Ba 2014) optimizer with a weight decay of 10 5. For CUB-200 and Cars196, a learning rate and the size of mini-batch are set to 5 10 5 and 128. For SOP and In-Shop, a learning rate and the size of mini-batch are set to 10 4 and 256. We mainly apply our method to Triplet loss (Schroff, Kalenichenko, and Philbin 2015). As a triplet sampling method, we employ the distance weighted sampling (Wu et al. 2017). The margin m of Triplet loss is set to 0.2. We summarized the hyperparameters of MDR: the configuration of the levels is initialized to three levels of { 3, 0, 3}, and the momentum γ is set to 0.9. λ is set differently for each dataset: 0.6 for CUB-200, 0.2 for Cars-196 and 0.1 for SOP and In-Shop. |