Hierarchical Occlusion Inference for Unseen Object Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object sensing technologies struggle to accurately detect and segment unseen object instances in cluttered scenes, particularly when objects are occluded by others, and require task-specific data sets and model retraining for new objects.
Innovation Solution
A hierarchical occlusion model that sequentially infers bounding boxes, visible masks, and occlusions of object instances using a computer system, employing image and object region of interest features, and fusing these through hierarchical fusion modules and fully connected layers to derive occlusion masks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If amodalinstance segmentation studies use trained object sets to segment amodal masks, then segmentation accuracy for known objects is improved, but the ability to detect new unseen objects deteriorates
Solution Approach 1:
The patent applies universality by developing a hierarchical occlusion model that can detect both known and unseen object instances through a unified framework. The model uses occlusion inference to identify object boundaries and visibility states without requiring category-specific training, enabling a single model to handle diverse object types including previously unknown categories.
Solution Approach 2:
The patent segments the object detection task into hierarchical levels: first inferring occlusion states and bounding boxes, then segmenting visible and amodal regions. This segmentation approach allows the system to handle unknown objects by first determining their spatial presence and occlusion status before performing category-specific analysis, thereby improving generalization to new object categories.
2Device complexity
If unseen object instance segmentation focuses on visible area recognition, then processing simplicity is improved, but the ability to detect occluded objects deteriorates
Solution Approach 1:
The patent applies preliminary action by first inferring occlusion states and generating amodal masks before performing visible region segmentation. The hierarchical model preliminarily determines which regions are occluded and creates corresponding amodal masks, then uses these to guide the segmentation of visible areas. This preliminary occlusion inference enables accurate detection of occluded objects while maintaining processing efficiency through a structured approach.
3Measurement precision
If object detection requires separate training data and model re-education for new objects, then detection accuracy for specific categories is improved, but the time and resources required for deployment deteriorates
Solution Approach 1:
The patent implements universality through a hierarchical occlusion model that detects objects across categories using a unified training approach. The model learns general occlusion inference patterns and spatial reasoning capabilities that apply to both known and unseen object categories, eliminating the need for separate training data and model re-education for each new category while maintaining high detection accuracy.
Data Source
AI summary
The hierarchical occlusion inference method according to the exemplary embodiment of the present disclosure includes: deriving a bounding box feature of the object instance by receiving a region of interest color-depth FPN feature and the object region of interest feature derived from a cluttered scene image including at least one object instance, deriving a visible feature of the object instance by fusing the object region of interest feature and the bounding box feature, deriving an amodal feature of the object instance by fusing the object region of interest feature, the bounding box feature, and the visible feature, deriving an occlusion feature of the object instance by fusing the object region of interest feature, the bounding box feature, the visible feature, and the amodal feature, and inferring occlusion of an object instance by de-convoluting the occlusion feature of the object instance.