Occluding Contour Detection via Dense Upsampling Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semantic segmentation systems face challenges in accurately detecting occluded objects in traffic environments due to the 'gridding issue' caused by standard dilated convolution operations and the loss of information during downsampling, which affects object contour detection and occlusion reasoning in autonomous driving systems.
Innovation Solution
The implementation of Dense Upsampling Convolution (DUC) and Hybrid Dilated Convolution (HDC) frameworks within a fully convolutional network (FCN) for pixel-level semantic segmentation, which enhances object contour detection by alleviating the gridding issue and recovering fine-detailed information, and formulates occluding contour detection as an image labeling task to identify occluded objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If standard dilated convolution operation is used to enlarge field of view, then context aggregation is improved, but gridding issue is produced causing loss of fine-detailed information
Solution Approach 1:
The patent segments the convolution operation into multiple stages with different dilation rates. Instead of using a single dilated convolution layer, the system applies serial dilated convolution layers with increasing dilation rates (1, 2, 4, 8) to progressively aggregate context at different scales while preserving fine details through the segmented multi-rate approach.
Solution Approach 2:
The patent applies different dilation rates to different spatial locations and scales. By using serial dilated convolution layers with varying rates, the system provides local quality adaptation where early layers preserve fine-grained local details with lower dilation rates, while later layers capture broader contextual information with higher dilation rates.
2Productivity
If max-pooling or strided convolution is used in CNN layers, then downsampling is achieved for feature extraction, but information loss occurs affecting object contour detection
Solution Approach 1:
The patent applies preliminary action by using deconvolutional layers before final contour detection to recover and reconstruct fine-detailed information that was lost during downsampling. The deconvolutional layers perform preliminary information recovery by upsampling feature maps and restoring spatial details before the final segmentation and contour detection stages.
Solution Approach 2:
The patent uses deconvolutional layers that act as mirrored versions of convolutional layers, copying the feature extraction process in reverse to reconstruct the original spatial information. By storing and utilizing pooling switches from the pooling step, the system creates a copied reconstruction path that recovers lost contour information.
3Reliability
If bounding box merging process is used in conventional object detection, then false positive rate is reduced, but occluded object detection fails when nearby bounding boxes are merged
Solution Approach 1:
The patent replaces the mechanical bounding box merging process with a neural network-based semantic segmentation approach. Instead of using rigid geometric operations to merge bounding boxes, the system uses learned pixel-wise classification with dilated convolutions and deconvolutions to detect object contours and instances, allowing the network to learn appropriate merging and separation decisions based on contextual information.
Solution Approach 2:
The patent changes the parameter space from discrete bounding box coordinates to continuous pixel-level probability maps. By transforming the detection problem into a pixel-wise classification task with soft probability outputs, the system can detect occluded objects through gradual probability transitions rather than hard bounding box boundaries, improving detection of partially occluded instances.
Data Source
AI summary
A system method for occluding contour detection using a fully convolutional neural network is disclosed. A particular embodiment includes: receiving an input image; producing a feature map from the input image by semantic segmentation; applying a Dense Upsampling Convolution (DUC) operation on the feature map to produce contour information of objects and object instances detected in the input image; and applying the contour information onto the input image.