Active Learning for Object Detection With Vectorized Dual Pseudo Loss and Multiple Instance Offset Constraint.

Area of Science:

• Computer Science
• Machine Learning
• Computer Vision

Background:

• Active learning methods for object detection struggle with missing ground truth labels for regression loss.
• Challenges include poor representation of unlabeled instance information and quality discrepancies between image and anchor levels.

Purpose of the Study:

• To propose an advanced active learning method for object detection.
• To address limitations in existing methods regarding regression loss, unlabeled data representation, and information quality.

Main Methods:

• A two-stage framework is introduced: image information quality evaluation and diversity-driven sampling.
• Pioneers a dual pseudo loss formulation for theoretically grounded regression loss estimation.
• Employs instance-level cosine similarity for effective removal of redundant images.

Main Results:

• The proposed method significantly outperforms state-of-the-art active learning approaches on PASCAL VOC and MS COCO datasets.
• The dual pseudo regression loss effectively captures regression information quality.
• Demonstrates robust performance in active learning for object detection tasks.

Conclusions:

• The novel active learning method enhances object detection by improving data selection strategies.
• The dual pseudo loss formulation is a key innovation for robust regression information quality assessment.
• The method offers a significant advancement for efficient and effective object detection model training.