Bi-directional Feature Reconstruction Network for Fine-Grained Few-Shot Image Classification
Authors: Jijie Wu, Dongliang Chang, Aneeshan Sain, Xiaoxu Li, Zhanyu Ma, Jie Cao, Jun Guo, Yi-Zhe Song
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on three widely used fine-grained image classification datasets consistently show considerable improvements compared with other methods. |
| Researcher Affiliation | Academia | Jijie Wu1, Dongliang Chang2, Aneeshan Sain3, Xiaoxu Li1*, Zhanyu Ma2, Jie Cao1, Jun Guo2, and Yi-Zhe Song3 1School of Computer and Communications, Lanzhou University of Technology, Lanzhou, China 2School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China 3Sketch X, CVSSP, University of Surrey, United Kingdom |
| Pseudocode | No | The paper describes the methodology using text and equations but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Codes are available at: https://github.com/PRIS-CV/Bi-FRN. |
| Open Datasets | Yes | To evaluate the performance of the proposed method, we selected three benchmark fine-grained datasets, CUB-200-2011 (CUB) (Wah et al. 2011) is a classic finegrained image classification dataset. ... Stanford-Dogs (Dogs) (Khosla et al. 2011) is a challenging fine-grained image categorization dataset. ... Stanford-Cars (Cars) (Krause et al. 2013) is also a commonly used benchmark dataset for fine-grained image classification. |
| Dataset Splits | Yes | For each dataset, we divided it into Dtrain, Dval and Dtest. The ratio of Dtrain, Dval and Dtest is same as the literature (Zhu, Liu, and Jiang 2020), and all images are resized to 84 84. ... The optimal model is selected by evaluating performance of the few-shot classification model on multiple tasks on Dval. ... Furthermore, for Res Net-12 models we train our proposed model using 15-way 5-shot episodes... ...we select the best-performing model based on the validation set, and validate every 20 epochs. |
| Hardware Specification | Yes | We implemented all our experiments on NVIDIA 3090Ti GPUs via Pytorch (Paszke et al. 2019). |
| Software Dependencies | No | The paper mentions 'Pytorch (Paszke et al. 2019)' but does not specify a version number for Pytorch or any other software dependencies. |
| Experiment Setup | Yes | In our experiments, we train all Conv-4 and Res Net-12 models for 1, 200 epochs using SGD with Nesterov momentum of 0.9. The initial learning rate is set to 0.1 and weight decay to 5e-4. Learning rate is decreased by a scaling factor of 10 after every 400 epochs. For Conv-4 models, we train the proposed models using 30-way 5-shot episodes, and test for 1-shot and 5-shot episodes. We use 15 query images per class in both settings. Furthermore, for Res Net-12 models we train our proposed model using 15-way 5-shot episodes, in order to save memory. |