Occluding Contour Detection Using DUC and HDC for Traffic Scenes
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 occluded object identification.
Innovation Solution
The implementation of Dense Upsampling Convolution (DUC) and Hybrid Dilation Convolution (HDC) frameworks within a fully convolutional network (FCN) framework, which enables pixel-level prediction and alleviates the 'gridding issue by using a range of dilation rates, effectively capturing detailed information and improving object contour detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If standard dilated convolution operations are used to enlarge the field of convolutional kernels, then context aggregation capability is improved, but a gridding issue is produced that degrades detection accuracy
Solution Approach 1:
The patent changes the parameter of dilation rate from fixed to variable, dynamically adjusting it based on the feature map depth. This allows the system to maintain high context aggregation capability while avoiding the gridding issue that occurs with standard fixed dilation rates, thereby preserving detection accuracy.
Solution Approach 2:
The patent introduces dynamic dilation rates that adapt according to the network depth and feature map characteristics. This dynamic adjustment enables the system to optimize context aggregation at different stages of the network while preventing the degradation of detection accuracy caused by static gridding patterns.
2Productivity
If max-pooling or strided convolution operations are applied in convolutional neural networks, then computational efficiency is improved, but information loss occurs that affects object contour detection
Solution Approach 1:
The patent introduces an occluding contour detection module as an intermediary that operates on feature maps at multiple scales. This module recovers lost contour information by detecting occluding contours in the semantic segmentation map, effectively compensating for information lost during downsampling while maintaining computational efficiency.
Solution Approach 2:
The patent segments the feature processing into multiple pathways: standard downsampling for efficiency, and a parallel occluding contour detection pathway for information recovery. This segmentation allows the system to maintain computational efficiency while recovering critical contour information that would otherwise be lost.
3Reliability
If bounding box merging process is used in conventional object detection frameworks, then false positive rate is reduced, but occluded objects are missed especially when the occluded region is large
Solution Approach 1:
The patent applies local quality by treating different regions of the image differently: visible regions use conventional bounding box merging for low false positive rates, while occluded regions use occluding contour detection for high detection accuracy. This localized approach allows the system to optimize for different objectives in different spatial regions.
Solution Approach 2:
The patent transitions from two-dimensional bounding box representations to three-dimensional occluding contour representations that include depth and occlusion information. This dimensional enhancement allows the system to distinguish between overlapping objects and occluded objects, improving occluded object detection while maintaining low false positive rates.
Data Source
AI summary
A system and 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; learning an array of upscaling filters to upscale the feature map into a final dense feature map of a desired size; applying the array of upscaling filters to 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.