Extreme R-CNN: Few-Shot Object Detection via Sample Synthesis and Knowledge Distillation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional object detectors need extensive annotations.
• Few-shot object detectors struggle with base category bias and sample variations.
• Limited data for unknown classes poses a significant challenge in object detection.

Purpose of the Study:

• To develop an effective few-shot object detection method using extremely limited original samples.
• To mitigate biases and improve robustness in object detection for novel categories.
• To introduce a Two-Stage Fine-Tuning Approach (TFA) named Extreme R-CNN.

Main Methods:

• Implemented sample synthesis via instance clipping and data augmentation.
• Enhanced Faster R-CNN by decoupling regression and classification components of the Region of Interest (RoI).
• Utilized knowledge distillation to transfer learning from base categories to novel ones.

Main Results:

• Achieved significant improvements over baseline methods on Microsoft COCO and PASCAL VOC datasets.
• Enhanced average precision for novel categories by up to 15% on PASCAL VOC and 6.1% on COCO.
• In a 1-shot scenario, model AP50 exceeded the baseline's 10-shot AP50 on PASCAL VOC.

Conclusions:

• The proposed Extreme R-CNN effectively addresses challenges in few-shot object detection with limited data.
• The TFA, integrating sample synthesis and knowledge distillation, demonstrates superior performance for novel categories.
• The method shows remarkable efficacy, outperforming baseline models even in low-shot scenarios.