Foreground Occupancy Map for Object Detection False Positives
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep object detectors trained with manually annotated images often suffer from 'background visual content leaking' (BVCL), where non-target visual content within bounding boxes leads to false detections, degrading detection performance in real-life scenarios with occluded or non-rectangular object contours.
Innovation Solution
A system that evaluates the impact of BVCL by using a foreground occupancy map to identify frequently occupied image regions, corrects the object detector through a modified training process by increasing the loss component for false detection areas, and generates a modified object detector that reduces false detections by accounting for background contributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manually annotated bounding boxes are used for training, then the training process is simple and fast, but background visual content leaks into the training data causing false detections
Solution Approach 1:
The system pre-computes foreground occupancy maps from training images before the actual object detection training. These maps identify regions that frequently contain foreground objects, which are then used to weight the loss function during training. This preliminary preparation allows the model to learn from bounding box annotations while being guided to focus on actual object regions rather than background areas, thereby maintaining training efficiency while improving detection accuracy.
2Ease of manufacture
If bounding boxes include all visual content within rectangular boundaries, then annotation is easier, but non-target objects are included causing detection errors
Solution Approach 1:
The system applies different weights to different spatial regions within bounding boxes based on the foreground occupancy map. Regions identified as frequently containing foreground objects receive higher weights, while background regions receive lower weights. This local differentiation allows the model to focus learning attention on relevant object areas while still using simple rectangular bounding boxes for annotation, thereby maintaining annotation simplicity while improving detection precision.
3Adaptability or versatility
If the object detector is trained to recognize all objects in bounding boxes, then comprehensive detection is achieved, but false positives increase due to background content
Solution Approach 1:
The system uses foreground occupancy maps as feedback to guide the training process. The occupancy information provides spatial priors about where foreground objects are likely to appear, which are incorporated into the loss function to penalize false detections in background regions. This feedback mechanism allows the model to maintain comprehensive detection coverage while reducing false positives by learning to distinguish between foreground and background regions.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for obtaining a foreground occupancy map for a camera view. The methods, systems, and apparatus include actions of determining an area of an image in which there is a false detection of an object, determining a likely contribution of the area to the false detection based on the foreground occupancy map, generating a modified object detector based on the likely contribution of the area, and detecting an object using the modified object detector.