Few-Shot Object Detection via Association and DIscrimination

Authors: Yuhang Cao, Jiaqi Wang, Ying Jin, Tong Wu, Kai Chen, Ziwei Liu, Dahua Lin

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on standard Pascal VOC and MS-COCO datasets demonstrate that FADI achieves new state-of-the-art performance, significantly improving the baseline in any shot/split by +18.7.
Researcher Affiliation Collaboration 1CUHK-Sense Time Joint Lab, The Chinese University of Hong Kong 2Shanghai AI Laboratory 3Sense Time Research 4S-Lab, Nanyang Technological University
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No In the ethics review checklist, the authors state: "Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No] We will release the code to ensure strict reproducibility upon the paper accepted."
Open Datasets Yes We conduct experiments on both PASCAL VOC (07 + 12) [7] and MS COCO [18] datasets.
Dataset Splits Yes In few-shot detection, the training set is composed of a base set DB = {x B i , y B i } with abundant data of classes CB, and a novel set DN = {x N i , y N i } with few-shot data of classes CN, where xi and yi indicate training samples and labels, respectively. The number of objects for each class in CN is K for K-shot detection. The model is expected to detect objects in the test set with classes in CB CN. ... We then construct a balanced training set with K shots per class.
Hardware Specification Yes In the ethics review checklist, the authors state: "Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Appendix 2."
Software Dependencies No We implement our methods based on MMDetection [3]. Faster-RCNN [21] with Feature Pyramid Network [17] and Res Net-101 [12] are adopted as base model. While MMDetection is mentioned, specific version numbers for it or other software dependencies are not provided.
Experiment Setup Yes Detailed settings are described in the supplementary material. ... During training, we freeze all parameters except the second linear layer FC 2, which means g( ; WN asso) is a single fc layer. We then construct a balanced training set with K shots per class.