Learning an Invariant and Equivariant Network for Weakly Supervised Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Weakly Supervised Object Detection (WSOD) is crucial for computer vision applications due to its reduced manual annotation costs.
• Current WSOD methods often suffer from instability and incomplete detection owing to indefinite and quality-agnostic frameworks.
• Issues stem from inconsistent learning of object variations and lack of localization quality awareness.

Purpose of the Study:

• To address the limitations of existing WSOD approaches.
• To develop a novel end-to-end network for more stable and comprehensive object detection.
• To enhance feature learning for WSOD without requiring additional manual annotations.

Main Methods:

• Introduction of the Invariant and Equivariant Network (IENet), an end-to-end deep learning model.
• Implementation of a flexible multi-branch online refinement strategy for comprehensive object perception.
• Utilizing progressive label propagation for affine-invariant learning and incorporating a rotation-equivariant branch for localization quality awareness.

Main Results:

• IENet achieves affine-invariant learning by propagating labels to transformed instances.
• The network incorporates rotation-equivariant learning for improved localization quality awareness.
• Significant performance boosts in WSOD on challenging natural and aerial scene datasets, establishing new state-of-the-art results.

Conclusions:

• IENet provides consistent and holistic feature learning for WSOD.
• The proposed methods effectively overcome the limitations of unstable and incomplete object detectors.
• The approach demonstrates superior performance on benchmark datasets, advancing the field of weakly supervised object detection.