Narrowing the Gap: Improved Detector Training With Noisy Location Annotations.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep learning models, particularly for object detection, rely heavily on large, accurately annotated datasets.
• Manual annotation, especially pixel-wise bounding boxes, is labor-intensive, time-consuming, and prone to human error, introducing noise.
• Noisy annotations can significantly degrade the performance of object detection algorithms.

Purpose of the Study:

• To investigate the impact of noisy location annotations on object detection performance.
• To develop a method for reducing the adverse effects of noisy annotations on user-side.
• To improve the robustness of object detection models against annotation inaccuracies.

Main Methods:

• Introduced synthesized noise into bounding box annotations to quantify performance degradation.
• Proposed a self-correction technique utilizing a Bayesian filter for prediction ensembles.
• Employed a Teacher-Student learning paradigm to leverage noisy annotations effectively.

Main Results:

• Significant performance drops were observed in both one-stage (e.g., FCOS) and two-stage (e.g., Faster R-CNN) detectors with noisy annotations.
• Synthesized noise reduced FCOS performance from 38.9% AP to 33.6% AP and Faster R-CNN from 37.8% AP to 33.7% AP on the COCO dataset.
• The proposed self-correction method successfully recovered degraded performance, increasing FCOS AP from 33.6% to 35.6%.

Conclusions:

• Noisy location annotations pose a substantial challenge to object detection model accuracy.
• The developed self-correction technique effectively mitigates the negative impact of noisy annotations.
• This approach offers a practical solution for improving object detection performance in real-world scenarios with imperfect data.