Related Experiment Video
Updated: Jul 15, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
Dual-Thresholded Heatmap-Guided Proposal Clustering and Negative Certainty Supervision with Enhanced Base Network for
Summary
The DANCE method improves weakly supervised object detection (WSOD) by generating better object proposals and incorporating background information. This approach enhances accuracy and speeds up model convergence for more efficient object detection.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Weakly supervised object detection (WSOD) methods reduce annotation costs but face challenges with inaccurate bounding box generation and semantic gaps.
- Existing WSOD approaches often produce pseudo ground truth (GT) boxes that are either incomplete or fail to differentiate adjacent instances.
- Limitations include a lack of background representation and slow convergence due to discarded proposals during optimization.
Purpose of the Study:
- To introduce a novel method, DANCE (Dual-thresholded heAtmapguided proposal clustering and Negative Certainty supervision with Enhanced base network), to address key limitations in WSOD.
- To improve the quality of pseudo GT boxes and enhance the representation of background information within detection models.
- To accelerate the convergence of WSOD models by effectively utilizing ignored proposals.
Main Methods:
- Development of a heatmap-guided proposal selector (HGPS) using dual thresholds for precise object proposal generation.
- Construction of a weakly supervised basic detection network (WSBDN) incorporating background class representation and heatmap pre-supervision.
- Implementation of a negative certainty supervision (NCS) loss applied to ignored proposals to speed up training.
Main Results:
- The proposed DANCE method significantly improves the accuracy of pseudo GT box generation, capturing full object extents and distinguishing adjacent instances.
- Experiments on PASCAL VOC and MS COCO datasets demonstrate the effectiveness and superiority of DANCE compared to existing state-of-the-art methods.
- The NCS loss effectively accelerates model convergence, addressing the issue of slow training in previous approaches.
Conclusions:
- DANCE offers a robust solution to critical challenges in weakly supervised object detection, enhancing both accuracy and efficiency.
- The method's ability to generate better object proposals and leverage background information marks a significant advancement in the field.
- The publicly available code facilitates further research and application of the DANCE method in computer vision tasks.