Cascade Meta-RCNN for Few-shot Object Detection
Shuting Li, Qian Jiang, Xin Jin, Nanqing Liu, Shiyu Chen, Shin‐Jye Lee · 2023
Data annotation is often a labor-intensive and time-consuming task, and in some cases, it is even challenging to collect certain data. Few-shot object detection (FSOD) addresses this challenge by recognizing objects with very few examples. Meta-learning-based FSOD usually trains models on a limited set of samples, enabling them to learn and generalize to new samples. Currently, meta-learning is widely utilized in the field of two-stage object detection, which aggregates query features and support features to obtain the final classification scores. Hence, there is a higher demand for accurate regions of interest (RoI). However, these methods only consider training samples with a single Intersection over Union (IoU) threshold. In this paper, we incorporate cascade structure into the existing Meta-RCNN model named Cascade Meta-RCNN. Specifically, the query feature is aggregated with prototypes of the support set. Then, the aggregated features are sequentially input into different RoI-Heads, which are trained with progressively increasing IoU thresholds. This method compels the network to generate more accurate query RoI features for matching with support prototypes. Additionally, we integrated the Channel and Spatial Attention (CSA) module into the model’s backbone, enhancing the network’s discriminative ability and further boosting its performance. To validate the effectiveness of our approach, we conducted a series of experiments on the PASCAL VOC dataset. The results demonstrate that our method outperforms the current state-of-the-art methods.