Related Experiment Video
Updated: Aug 4, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
592
Selecting High-Quality Proposals for Weakly Supervised Object Detection With Bottom-Up Aggregated Attention and
Summary
This study introduces novel methods to improve weakly supervised object detection (WSOD) by enhancing proposal selection. The techniques address imprecise supervision, leading to state-of-the-art performance in object detection tasks.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Weakly supervised object detection (WSOD) trains models using only image-category labels, reducing annotation costs.
- Existing WSOD methods often suffer from imprecise supervision, leading to issues like part domination and untight bounding boxes.
- Two-phase frameworks, involving instance mining and refinement, are common but limited by the quality of initial supervision.
Purpose of the Study:
- To enhance the quality of proposals used for supervision in weakly supervised object detection.
- To mitigate the adverse effects of imprecise supervision, specifically part domination and untight bounding boxes.
- To achieve state-of-the-art results in WSOD by improving the accuracy of object localization and classification.
Main Methods:
- Proposed a bottom-up aggregated attention mechanism to improve the location representation of object proposals by incorporating low-level features.
- Introduced a phase-aware loss function to measure supervision quality during the instance mining phase, highlighting correct boxes and suppressing inaccurate ones.
- Integrated these modules into an online instance classifier refinement framework for WSOD.
Main Results:
- The bottom-up aggregated attention module effectively improves the scoring of proposals corresponding to entire objects.
- The phase-aware loss successfully distinguishes between correct and untight bounding boxes, refining supervision quality.
- The unified framework demonstrated significant performance improvements on benchmark datasets like PASCAL VOC and MS COCO.
Conclusions:
- The proposed methods effectively address key limitations in weakly supervised object detection, particularly imprecise supervision.
- The novel attention mechanism and loss function contribute to more accurate object localization and classification.
- The approach achieves state-of-the-art performance, offering a robust solution for WSOD.
