PCL: Proposal Cluster Learning for Weakly Supervised Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Weakly Supervised Object Detection (WSOD) is crucial for object recognition, utilizing only image-level annotations.
• Existing WSOD methods often transform the problem into image classification via Multiple Instance Learning (MIL).

Purpose of the Study:

• To propose a novel deep network for WSOD that overcomes limitations of MIL-based approaches.
• To enhance object detection by refining instance classifiers through an iterative process using proposal clusters.

Main Methods:

• A novel deep network architecture is proposed for WSOD.
• The method generates proposal clusters of spatially adjacent, object-associated proposals.
• Iterative refinement of instance classifiers is performed online using multiple CNN streams, starting with an MIL network.

Main Results:

• The proposed method demonstrates superior performance on PASCAL VOC, ImageNet detection, and MS-COCO benchmarks.
• Results significantly outperform previous state-of-the-art WSOD techniques.
• The strategy of refining instance classifiers using proposal clusters proves effective in preventing focus on object parts.

Conclusions:

• The novel deep network effectively addresses challenges in WSOD.
• Iterative instance classifier refinement using proposal clusters offers a more robust approach than direct label assignment.
• The method achieves significant performance gains in weakly supervised object detection.