Meta Faster R-CNN: Towards Accurate Few-Shot Object Detection with Attentive Feature Alignment

Authors: Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, Shih-Fu Chang780-789

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental Results Datasets We use two widely-used FSOD benchmarks MSCOCO (Lin et al. 2014) and PASCAL VOC (Everingham et al. 2010) for model evaluation, and follow FSOD settings the same as previous works (Kang et al. 2019; Wang et al. 2020) by using the exact same few-shot images for fair comparison. More implementation details are included in the supplementary material. Ablation Study Effectiveness of our Meta-RPN. We compare three different proposal generation methods for novel classes (RPN, Attention-RPN (Fan et al. 2020), and our Meta-RPN) in Table 1 (b), (c) and (d) and Table 2.
Researcher Affiliation Academia Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, Shih-Fu Chang Columbia University {gh2561, sh3813, jiawei.m, yh3330, sc250}@columbia.edu
Pseudocode No The paper describes the proposed methods using text and diagrams but does not include any formal pseudocode or algorithm blocks.
Open Source Code No The authors report these results at https://github.com/Young XIAO13/Few Shot Detection. This link refers to results reported by other authors, not the source code for the methodology described in this paper.
Open Datasets Yes We use two widely-used FSOD benchmarks MSCOCO (Lin et al. 2014) and PASCAL VOC (Everingham et al. 2010) for model evaluation
Dataset Splits Yes We make sure that the total number instances for each novel class is exactly k-shot in the sampled dataset, which are also used as the support set during testing.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU models, CPU types, or cloud instance specifications) used for running the experiments.
Software Dependencies No The paper mentions architectural components like "Res Net-50/101", "Faster R-CNN", and "Ro IAlign", but does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA 11.x).
Experiment Setup No The paper describes the training framework and mentions the use of binary cross-entropy loss and smooth L1 loss, and a 1:3 ratio for positive and negative matching pairs. However, it does not provide specific hyperparameter values such as learning rate, batch size, number of epochs, or optimizer details.